xref: /linux/fs/nfs/pnfs.c (revision a77f02e8489673cc80948a6f30ed262db1dee8d6)
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 const struct pnfs_layoutdriver_type *pnfs_find_layoutdriver(u32 id)
96 {
97 	return find_pnfs_driver(id);
98 }
99 
100 void pnfs_put_layoutdriver(const struct pnfs_layoutdriver_type *ld)
101 {
102 	if (ld)
103 		module_put(ld->owner);
104 }
105 
106 void
107 unset_pnfs_layoutdriver(struct nfs_server *nfss)
108 {
109 	if (nfss->pnfs_curr_ld) {
110 		if (nfss->pnfs_curr_ld->clear_layoutdriver)
111 			nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
112 		/* Decrement the MDS count. Purge the deviceid cache if zero */
113 		if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
114 			nfs4_deviceid_purge_client(nfss->nfs_client);
115 		module_put(nfss->pnfs_curr_ld->owner);
116 	}
117 	nfss->pnfs_curr_ld = NULL;
118 }
119 
120 /*
121  * When the server sends a list of layout types, we choose one in the order
122  * given in the list below.
123  *
124  * FIXME: should this list be configurable in some fashion? module param?
125  * 	  mount option? something else?
126  */
127 static const u32 ld_prefs[] = {
128 	LAYOUT_SCSI,
129 	LAYOUT_BLOCK_VOLUME,
130 	LAYOUT_OSD2_OBJECTS,
131 	LAYOUT_FLEX_FILES,
132 	LAYOUT_NFSV4_1_FILES,
133 	0
134 };
135 
136 static int
137 ld_cmp(const void *e1, const void *e2)
138 {
139 	u32 ld1 = *((u32 *)e1);
140 	u32 ld2 = *((u32 *)e2);
141 	int i;
142 
143 	for (i = 0; ld_prefs[i] != 0; i++) {
144 		if (ld1 == ld_prefs[i])
145 			return -1;
146 
147 		if (ld2 == ld_prefs[i])
148 			return 1;
149 	}
150 	return 0;
151 }
152 
153 /*
154  * Try to set the server's pnfs module to the pnfs layout type specified by id.
155  * Currently only one pNFS layout driver per filesystem is supported.
156  *
157  * @ids array of layout types supported by MDS.
158  */
159 void
160 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
161 		      struct nfs_fsinfo *fsinfo)
162 {
163 	struct pnfs_layoutdriver_type *ld_type = NULL;
164 	u32 id;
165 	int i;
166 
167 	if (fsinfo->nlayouttypes == 0)
168 		goto out_no_driver;
169 	if (!(server->nfs_client->cl_exchange_flags &
170 		 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
171 		printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
172 			__func__, server->nfs_client->cl_exchange_flags);
173 		goto out_no_driver;
174 	}
175 
176 	sort(fsinfo->layouttype, fsinfo->nlayouttypes,
177 		sizeof(*fsinfo->layouttype), ld_cmp, NULL);
178 
179 	for (i = 0; i < fsinfo->nlayouttypes; i++) {
180 		id = fsinfo->layouttype[i];
181 		ld_type = find_pnfs_driver(id);
182 		if (!ld_type) {
183 			request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
184 					id);
185 			ld_type = find_pnfs_driver(id);
186 		}
187 		if (ld_type)
188 			break;
189 	}
190 
191 	if (!ld_type) {
192 		dprintk("%s: No pNFS module found!\n", __func__);
193 		goto out_no_driver;
194 	}
195 
196 	server->pnfs_curr_ld = ld_type;
197 	if (ld_type->set_layoutdriver
198 	    && ld_type->set_layoutdriver(server, mntfh)) {
199 		printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
200 			"driver %u.\n", __func__, id);
201 		module_put(ld_type->owner);
202 		goto out_no_driver;
203 	}
204 	/* Bump the MDS count */
205 	atomic_inc(&server->nfs_client->cl_mds_count);
206 
207 	dprintk("%s: pNFS module for %u set\n", __func__, id);
208 	return;
209 
210 out_no_driver:
211 	dprintk("%s: Using NFSv4 I/O\n", __func__);
212 	server->pnfs_curr_ld = NULL;
213 }
214 
215 int
216 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
217 {
218 	int status = -EINVAL;
219 	struct pnfs_layoutdriver_type *tmp;
220 
221 	if (ld_type->id == 0) {
222 		printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
223 		return status;
224 	}
225 	if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
226 		printk(KERN_ERR "NFS: %s Layout driver must provide "
227 		       "alloc_lseg and free_lseg.\n", __func__);
228 		return status;
229 	}
230 
231 	spin_lock(&pnfs_spinlock);
232 	tmp = find_pnfs_driver_locked(ld_type->id);
233 	if (!tmp) {
234 		list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
235 		status = 0;
236 		dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
237 			ld_type->name);
238 	} else {
239 		printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
240 			__func__, ld_type->id);
241 	}
242 	spin_unlock(&pnfs_spinlock);
243 
244 	return status;
245 }
246 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
247 
248 void
249 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
250 {
251 	dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
252 	spin_lock(&pnfs_spinlock);
253 	list_del(&ld_type->pnfs_tblid);
254 	spin_unlock(&pnfs_spinlock);
255 }
256 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
257 
258 /*
259  * pNFS client layout cache
260  */
261 
262 /* Need to hold i_lock if caller does not already hold reference */
263 void
264 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
265 {
266 	refcount_inc(&lo->plh_refcount);
267 }
268 
269 static struct pnfs_layout_hdr *
270 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
271 {
272 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
273 	return ld->alloc_layout_hdr(ino, gfp_flags);
274 }
275 
276 static void
277 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
278 {
279 	struct nfs_server *server = NFS_SERVER(lo->plh_inode);
280 	struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
281 
282 	if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
283 		struct nfs_client *clp = server->nfs_client;
284 
285 		spin_lock(&clp->cl_lock);
286 		list_del_rcu(&lo->plh_layouts);
287 		spin_unlock(&clp->cl_lock);
288 	}
289 	put_cred(lo->plh_lc_cred);
290 	return ld->free_layout_hdr(lo);
291 }
292 
293 static void
294 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
295 {
296 	struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
297 	dprintk("%s: freeing layout cache %p\n", __func__, lo);
298 	nfsi->layout = NULL;
299 	/* Reset MDS Threshold I/O counters */
300 	nfsi->write_io = 0;
301 	nfsi->read_io = 0;
302 }
303 
304 void
305 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
306 {
307 	struct inode *inode;
308 	unsigned long i_state;
309 
310 	if (!lo)
311 		return;
312 	inode = lo->plh_inode;
313 	pnfs_layoutreturn_before_put_layout_hdr(lo);
314 
315 	if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
316 		if (!list_empty(&lo->plh_segs))
317 			WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
318 		pnfs_detach_layout_hdr(lo);
319 		i_state = inode->i_state;
320 		spin_unlock(&inode->i_lock);
321 		pnfs_free_layout_hdr(lo);
322 		/* Notify pnfs_destroy_layout_final() that we're done */
323 		if (i_state & (I_FREEING | I_CLEAR))
324 			wake_up_var(lo);
325 	}
326 }
327 
328 static struct inode *
329 pnfs_grab_inode_layout_hdr(struct pnfs_layout_hdr *lo)
330 {
331 	struct inode *inode = igrab(lo->plh_inode);
332 	if (inode)
333 		return inode;
334 	set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
335 	return NULL;
336 }
337 
338 /*
339  * Compare 2 layout stateid sequence ids, to see which is newer,
340  * taking into account wraparound issues.
341  */
342 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
343 {
344 	return (s32)(s1 - s2) > 0;
345 }
346 
347 static void pnfs_barrier_update(struct pnfs_layout_hdr *lo, u32 newseq)
348 {
349 	if (pnfs_seqid_is_newer(newseq, lo->plh_barrier) || !lo->plh_barrier)
350 		lo->plh_barrier = newseq;
351 }
352 
353 static void
354 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
355 			 u32 seq)
356 {
357 	if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
358 		iomode = IOMODE_ANY;
359 	lo->plh_return_iomode = iomode;
360 	set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
361 	/*
362 	 * We must set lo->plh_return_seq to avoid livelocks with
363 	 * pnfs_layout_need_return()
364 	 */
365 	if (seq == 0)
366 		seq = be32_to_cpu(lo->plh_stateid.seqid);
367 	if (!lo->plh_return_seq || pnfs_seqid_is_newer(seq, lo->plh_return_seq))
368 		lo->plh_return_seq = seq;
369 	pnfs_barrier_update(lo, seq);
370 }
371 
372 static void
373 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
374 {
375 	struct pnfs_layout_segment *lseg;
376 	lo->plh_return_iomode = 0;
377 	lo->plh_return_seq = 0;
378 	clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
379 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
380 		if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
381 			continue;
382 		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
383 	}
384 }
385 
386 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
387 {
388 	clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
389 	clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
390 	smp_mb__after_atomic();
391 	wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
392 	rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
393 }
394 
395 static void
396 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
397 		struct list_head *free_me)
398 {
399 	clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
400 	clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
401 	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
402 		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
403 	if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
404 		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
405 }
406 
407 /*
408  * Update the seqid of a layout stateid after receiving
409  * NFS4ERR_OLD_STATEID
410  */
411 bool nfs4_layout_refresh_old_stateid(nfs4_stateid *dst,
412 		struct pnfs_layout_range *dst_range,
413 		struct inode *inode)
414 {
415 	struct pnfs_layout_hdr *lo;
416 	struct pnfs_layout_range range = {
417 		.iomode = IOMODE_ANY,
418 		.offset = 0,
419 		.length = NFS4_MAX_UINT64,
420 	};
421 	bool ret = false;
422 	LIST_HEAD(head);
423 	int err;
424 
425 	spin_lock(&inode->i_lock);
426 	lo = NFS_I(inode)->layout;
427 	if (lo &&  pnfs_layout_is_valid(lo) &&
428 	    nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
429 		/* Is our call using the most recent seqid? If so, bump it */
430 		if (!nfs4_stateid_is_newer(&lo->plh_stateid, dst)) {
431 			nfs4_stateid_seqid_inc(dst);
432 			ret = true;
433 			goto out;
434 		}
435 		/* Try to update the seqid to the most recent */
436 		err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
437 		if (err != -EBUSY) {
438 			dst->seqid = lo->plh_stateid.seqid;
439 			*dst_range = range;
440 			ret = true;
441 		}
442 	}
443 out:
444 	spin_unlock(&inode->i_lock);
445 	pnfs_free_lseg_list(&head);
446 	return ret;
447 }
448 
449 /*
450  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
451  *
452  * In order to continue using the pnfs_layout_hdr, a full recovery
453  * is required.
454  * Note that caller must hold inode->i_lock.
455  */
456 int
457 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
458 		struct list_head *lseg_list)
459 {
460 	struct pnfs_layout_range range = {
461 		.iomode = IOMODE_ANY,
462 		.offset = 0,
463 		.length = NFS4_MAX_UINT64,
464 	};
465 	struct pnfs_layout_segment *lseg, *next;
466 
467 	set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
468 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
469 		pnfs_clear_lseg_state(lseg, lseg_list);
470 	pnfs_clear_layoutreturn_info(lo);
471 	pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
472 	set_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags);
473 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
474 	    !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
475 		pnfs_clear_layoutreturn_waitbit(lo);
476 	return !list_empty(&lo->plh_segs);
477 }
478 
479 static int
480 pnfs_iomode_to_fail_bit(u32 iomode)
481 {
482 	return iomode == IOMODE_RW ?
483 		NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
484 }
485 
486 static void
487 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
488 {
489 	lo->plh_retry_timestamp = jiffies;
490 	if (!test_and_set_bit(fail_bit, &lo->plh_flags))
491 		refcount_inc(&lo->plh_refcount);
492 }
493 
494 static void
495 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
496 {
497 	if (test_and_clear_bit(fail_bit, &lo->plh_flags))
498 		refcount_dec(&lo->plh_refcount);
499 }
500 
501 static void
502 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
503 {
504 	struct inode *inode = lo->plh_inode;
505 	struct pnfs_layout_range range = {
506 		.iomode = iomode,
507 		.offset = 0,
508 		.length = NFS4_MAX_UINT64,
509 	};
510 	LIST_HEAD(head);
511 
512 	spin_lock(&inode->i_lock);
513 	pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
514 	pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
515 	spin_unlock(&inode->i_lock);
516 	pnfs_free_lseg_list(&head);
517 	dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
518 			iomode == IOMODE_RW ?  "RW" : "READ");
519 }
520 
521 static bool
522 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
523 {
524 	unsigned long start, end;
525 	int fail_bit = pnfs_iomode_to_fail_bit(iomode);
526 
527 	if (test_bit(fail_bit, &lo->plh_flags) == 0)
528 		return false;
529 	end = jiffies;
530 	start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
531 	if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
532 		/* It is time to retry the failed layoutgets */
533 		pnfs_layout_clear_fail_bit(lo, fail_bit);
534 		return false;
535 	}
536 	return true;
537 }
538 
539 static void
540 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
541 		const struct pnfs_layout_range *range,
542 		const nfs4_stateid *stateid)
543 {
544 	INIT_LIST_HEAD(&lseg->pls_list);
545 	INIT_LIST_HEAD(&lseg->pls_lc_list);
546 	INIT_LIST_HEAD(&lseg->pls_commits);
547 	refcount_set(&lseg->pls_refcount, 1);
548 	set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
549 	lseg->pls_layout = lo;
550 	lseg->pls_range = *range;
551 	lseg->pls_seq = be32_to_cpu(stateid->seqid);
552 }
553 
554 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
555 {
556 	if (lseg != NULL) {
557 		struct inode *inode = lseg->pls_layout->plh_inode;
558 		NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
559 	}
560 }
561 
562 static void
563 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
564 		struct pnfs_layout_segment *lseg)
565 {
566 	WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
567 	list_del_init(&lseg->pls_list);
568 	/* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
569 	refcount_dec(&lo->plh_refcount);
570 	if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
571 		return;
572 	if (list_empty(&lo->plh_segs) &&
573 	    !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
574 	    !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
575 		if (atomic_read(&lo->plh_outstanding) == 0)
576 			set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
577 		clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
578 	}
579 }
580 
581 static bool
582 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
583 		struct pnfs_layout_segment *lseg)
584 {
585 	if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
586 	    pnfs_layout_is_valid(lo)) {
587 		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
588 		list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
589 		return true;
590 	}
591 	return false;
592 }
593 
594 void
595 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
596 {
597 	struct pnfs_layout_hdr *lo;
598 	struct inode *inode;
599 
600 	if (!lseg)
601 		return;
602 
603 	dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
604 		refcount_read(&lseg->pls_refcount),
605 		test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
606 
607 	lo = lseg->pls_layout;
608 	inode = lo->plh_inode;
609 
610 	if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
611 		pnfs_get_layout_hdr(lo);
612 		pnfs_layout_remove_lseg(lo, lseg);
613 		if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
614 			lseg = NULL;
615 		spin_unlock(&inode->i_lock);
616 		pnfs_free_lseg(lseg);
617 		pnfs_put_layout_hdr(lo);
618 	}
619 }
620 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
621 
622 /*
623  * is l2 fully contained in l1?
624  *   start1                             end1
625  *   [----------------------------------)
626  *           start2           end2
627  *           [----------------)
628  */
629 static bool
630 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
631 		 const struct pnfs_layout_range *l2)
632 {
633 	u64 start1 = l1->offset;
634 	u64 end1 = pnfs_end_offset(start1, l1->length);
635 	u64 start2 = l2->offset;
636 	u64 end2 = pnfs_end_offset(start2, l2->length);
637 
638 	return (start1 <= start2) && (end1 >= end2);
639 }
640 
641 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
642 		struct list_head *tmp_list)
643 {
644 	if (!refcount_dec_and_test(&lseg->pls_refcount))
645 		return false;
646 	pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
647 	list_add(&lseg->pls_list, tmp_list);
648 	return true;
649 }
650 
651 /* Returns 1 if lseg is removed from list, 0 otherwise */
652 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
653 			     struct list_head *tmp_list)
654 {
655 	int rv = 0;
656 
657 	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
658 		/* Remove the reference keeping the lseg in the
659 		 * list.  It will now be removed when all
660 		 * outstanding io is finished.
661 		 */
662 		dprintk("%s: lseg %p ref %d\n", __func__, lseg,
663 			refcount_read(&lseg->pls_refcount));
664 		if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
665 			rv = 1;
666 	}
667 	return rv;
668 }
669 
670 static bool
671 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
672 		 const struct pnfs_layout_range *recall_range)
673 {
674 	return (recall_range->iomode == IOMODE_ANY ||
675 		lseg_range->iomode == recall_range->iomode) &&
676 	       pnfs_lseg_range_intersecting(lseg_range, recall_range);
677 }
678 
679 static bool
680 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
681 		const struct pnfs_layout_range *recall_range,
682 		u32 seq)
683 {
684 	if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
685 		return false;
686 	if (recall_range == NULL)
687 		return true;
688 	return pnfs_should_free_range(&lseg->pls_range, recall_range);
689 }
690 
691 /**
692  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
693  * @lo: layout header containing the lsegs
694  * @tmp_list: list head where doomed lsegs should go
695  * @recall_range: optional recall range argument to match (may be NULL)
696  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
697  *
698  * Walk the list of lsegs in the layout header, and tear down any that should
699  * be destroyed. If "recall_range" is specified then the segment must match
700  * that range. If "seq" is non-zero, then only match segments that were handed
701  * out at or before that sequence.
702  *
703  * Returns number of matching invalid lsegs remaining in list after scanning
704  * it and purging them.
705  */
706 int
707 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
708 			    struct list_head *tmp_list,
709 			    const struct pnfs_layout_range *recall_range,
710 			    u32 seq)
711 {
712 	struct pnfs_layout_segment *lseg, *next;
713 	struct nfs_server *server = NFS_SERVER(lo->plh_inode);
714 	int remaining = 0;
715 
716 	dprintk("%s:Begin lo %p\n", __func__, lo);
717 
718 	if (list_empty(&lo->plh_segs))
719 		return 0;
720 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
721 		if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
722 			dprintk("%s: freeing lseg %p iomode %d seq %u "
723 				"offset %llu length %llu\n", __func__,
724 				lseg, lseg->pls_range.iomode, lseg->pls_seq,
725 				lseg->pls_range.offset, lseg->pls_range.length);
726 			if (mark_lseg_invalid(lseg, tmp_list))
727 				continue;
728 			remaining++;
729 			pnfs_lseg_cancel_io(server, lseg);
730 		}
731 	dprintk("%s:Return %i\n", __func__, remaining);
732 	return remaining;
733 }
734 
735 static void
736 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
737 		struct list_head *free_me,
738 		const struct pnfs_layout_range *range,
739 		u32 seq)
740 {
741 	struct pnfs_layout_segment *lseg, *next;
742 
743 	list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
744 		if (pnfs_match_lseg_recall(lseg, range, seq))
745 			list_move_tail(&lseg->pls_list, free_me);
746 	}
747 }
748 
749 /* note free_me must contain lsegs from a single layout_hdr */
750 void
751 pnfs_free_lseg_list(struct list_head *free_me)
752 {
753 	struct pnfs_layout_segment *lseg, *tmp;
754 
755 	if (list_empty(free_me))
756 		return;
757 
758 	list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
759 		list_del(&lseg->pls_list);
760 		pnfs_free_lseg(lseg);
761 	}
762 }
763 
764 static struct pnfs_layout_hdr *__pnfs_destroy_layout(struct nfs_inode *nfsi)
765 {
766 	struct pnfs_layout_hdr *lo;
767 	LIST_HEAD(tmp_list);
768 
769 	spin_lock(&nfsi->vfs_inode.i_lock);
770 	lo = nfsi->layout;
771 	if (lo) {
772 		pnfs_get_layout_hdr(lo);
773 		pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
774 		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
775 		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
776 		spin_unlock(&nfsi->vfs_inode.i_lock);
777 		pnfs_free_lseg_list(&tmp_list);
778 		nfs_commit_inode(&nfsi->vfs_inode, 0);
779 		pnfs_put_layout_hdr(lo);
780 	} else
781 		spin_unlock(&nfsi->vfs_inode.i_lock);
782 	return lo;
783 }
784 
785 void pnfs_destroy_layout(struct nfs_inode *nfsi)
786 {
787 	__pnfs_destroy_layout(nfsi);
788 }
789 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
790 
791 static bool pnfs_layout_removed(struct nfs_inode *nfsi,
792 				struct pnfs_layout_hdr *lo)
793 {
794 	bool ret;
795 
796 	spin_lock(&nfsi->vfs_inode.i_lock);
797 	ret = nfsi->layout != lo;
798 	spin_unlock(&nfsi->vfs_inode.i_lock);
799 	return ret;
800 }
801 
802 void pnfs_destroy_layout_final(struct nfs_inode *nfsi)
803 {
804 	struct pnfs_layout_hdr *lo = __pnfs_destroy_layout(nfsi);
805 
806 	if (lo)
807 		wait_var_event(lo, pnfs_layout_removed(nfsi, lo));
808 }
809 
810 static bool
811 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
812 		struct list_head *layout_list)
813 {
814 	struct pnfs_layout_hdr *lo;
815 	bool ret = false;
816 
817 	spin_lock(&inode->i_lock);
818 	lo = NFS_I(inode)->layout;
819 	if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
820 		pnfs_get_layout_hdr(lo);
821 		list_add(&lo->plh_bulk_destroy, layout_list);
822 		ret = true;
823 	}
824 	spin_unlock(&inode->i_lock);
825 	return ret;
826 }
827 
828 /* Caller must hold rcu_read_lock and clp->cl_lock */
829 static int
830 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
831 		struct nfs_server *server,
832 		struct list_head *layout_list)
833 	__must_hold(&clp->cl_lock)
834 	__must_hold(RCU)
835 {
836 	struct pnfs_layout_hdr *lo, *next;
837 	struct inode *inode;
838 
839 	list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
840 		if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
841 		    test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
842 		    !list_empty(&lo->plh_bulk_destroy))
843 			continue;
844 		/* If the sb is being destroyed, just bail */
845 		if (!nfs_sb_active(server->super))
846 			break;
847 		inode = pnfs_grab_inode_layout_hdr(lo);
848 		if (inode != NULL) {
849 			if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags))
850 				list_del_rcu(&lo->plh_layouts);
851 			if (pnfs_layout_add_bulk_destroy_list(inode,
852 						layout_list))
853 				continue;
854 			rcu_read_unlock();
855 			spin_unlock(&clp->cl_lock);
856 			iput(inode);
857 		} else {
858 			rcu_read_unlock();
859 			spin_unlock(&clp->cl_lock);
860 		}
861 		nfs_sb_deactive(server->super);
862 		spin_lock(&clp->cl_lock);
863 		rcu_read_lock();
864 		return -EAGAIN;
865 	}
866 	return 0;
867 }
868 
869 static int
870 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
871 		bool is_bulk_recall)
872 {
873 	struct pnfs_layout_hdr *lo;
874 	struct inode *inode;
875 	LIST_HEAD(lseg_list);
876 	int ret = 0;
877 
878 	while (!list_empty(layout_list)) {
879 		lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
880 				plh_bulk_destroy);
881 		dprintk("%s freeing layout for inode %lu\n", __func__,
882 			lo->plh_inode->i_ino);
883 		inode = lo->plh_inode;
884 
885 		pnfs_layoutcommit_inode(inode, false);
886 
887 		spin_lock(&inode->i_lock);
888 		list_del_init(&lo->plh_bulk_destroy);
889 		if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
890 			if (is_bulk_recall)
891 				set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
892 			ret = -EAGAIN;
893 		}
894 		spin_unlock(&inode->i_lock);
895 		pnfs_free_lseg_list(&lseg_list);
896 		/* Free all lsegs that are attached to commit buckets */
897 		nfs_commit_inode(inode, 0);
898 		pnfs_put_layout_hdr(lo);
899 		nfs_iput_and_deactive(inode);
900 	}
901 	return ret;
902 }
903 
904 int
905 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
906 		struct nfs_fsid *fsid,
907 		bool is_recall)
908 {
909 	struct nfs_server *server;
910 	LIST_HEAD(layout_list);
911 
912 	spin_lock(&clp->cl_lock);
913 	rcu_read_lock();
914 restart:
915 	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
916 		if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
917 			continue;
918 		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
919 				server,
920 				&layout_list) != 0)
921 			goto restart;
922 	}
923 	rcu_read_unlock();
924 	spin_unlock(&clp->cl_lock);
925 
926 	if (list_empty(&layout_list))
927 		return 0;
928 	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
929 }
930 
931 int
932 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
933 		bool is_recall)
934 {
935 	struct nfs_server *server;
936 	LIST_HEAD(layout_list);
937 
938 	spin_lock(&clp->cl_lock);
939 	rcu_read_lock();
940 restart:
941 	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
942 		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
943 					server,
944 					&layout_list) != 0)
945 			goto restart;
946 	}
947 	rcu_read_unlock();
948 	spin_unlock(&clp->cl_lock);
949 
950 	if (list_empty(&layout_list))
951 		return 0;
952 	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
953 }
954 
955 /*
956  * Called by the state manager to remove all layouts established under an
957  * expired lease.
958  */
959 void
960 pnfs_destroy_all_layouts(struct nfs_client *clp)
961 {
962 	nfs4_deviceid_mark_client_invalid(clp);
963 	nfs4_deviceid_purge_client(clp);
964 
965 	pnfs_destroy_layouts_byclid(clp, false);
966 }
967 
968 static void
969 pnfs_set_layout_cred(struct pnfs_layout_hdr *lo, const struct cred *cred)
970 {
971 	const struct cred *old;
972 
973 	if (cred && cred_fscmp(lo->plh_lc_cred, cred) != 0) {
974 		old = xchg(&lo->plh_lc_cred, get_cred(cred));
975 		put_cred(old);
976 	}
977 }
978 
979 /* update lo->plh_stateid with new if is more recent */
980 void
981 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
982 			const struct cred *cred, bool update_barrier)
983 {
984 	u32 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
985 	u32 newseq = be32_to_cpu(new->seqid);
986 
987 	if (!pnfs_layout_is_valid(lo)) {
988 		pnfs_set_layout_cred(lo, cred);
989 		nfs4_stateid_copy(&lo->plh_stateid, new);
990 		lo->plh_barrier = newseq;
991 		pnfs_clear_layoutreturn_info(lo);
992 		clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
993 		return;
994 	}
995 
996 	if (pnfs_seqid_is_newer(newseq, oldseq))
997 		nfs4_stateid_copy(&lo->plh_stateid, new);
998 
999 	if (update_barrier) {
1000 		pnfs_barrier_update(lo, newseq);
1001 		return;
1002 	}
1003 	/*
1004 	 * Because of wraparound, we want to keep the barrier
1005 	 * "close" to the current seqids. We really only want to
1006 	 * get here from a layoutget call.
1007 	 */
1008 	if (atomic_read(&lo->plh_outstanding) == 1)
1009 		 pnfs_barrier_update(lo, be32_to_cpu(lo->plh_stateid.seqid));
1010 }
1011 
1012 static bool
1013 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
1014 		const nfs4_stateid *stateid)
1015 {
1016 	u32 seqid = be32_to_cpu(stateid->seqid);
1017 
1018 	return lo->plh_barrier && pnfs_seqid_is_newer(lo->plh_barrier, seqid);
1019 }
1020 
1021 /* lget is set to 1 if called from inside send_layoutget call chain */
1022 static bool
1023 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
1024 {
1025 	return lo->plh_block_lgets ||
1026 		test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
1027 }
1028 
1029 static struct nfs_server *
1030 pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
1031 {
1032 	struct nfs_server *server;
1033 
1034 	if (inode) {
1035 		server = NFS_SERVER(inode);
1036 	} else {
1037 		struct dentry *parent_dir = dget_parent(ctx->dentry);
1038 		server = NFS_SERVER(parent_dir->d_inode);
1039 		dput(parent_dir);
1040 	}
1041 	return server;
1042 }
1043 
1044 static void nfs4_free_pages(struct page **pages, size_t size)
1045 {
1046 	int i;
1047 
1048 	if (!pages)
1049 		return;
1050 
1051 	for (i = 0; i < size; i++) {
1052 		if (!pages[i])
1053 			break;
1054 		__free_page(pages[i]);
1055 	}
1056 	kfree(pages);
1057 }
1058 
1059 static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1060 {
1061 	struct page **pages;
1062 	int i;
1063 
1064 	pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
1065 	if (!pages) {
1066 		dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1067 		return NULL;
1068 	}
1069 
1070 	for (i = 0; i < size; i++) {
1071 		pages[i] = alloc_page(gfp_flags);
1072 		if (!pages[i]) {
1073 			dprintk("%s: failed to allocate page\n", __func__);
1074 			nfs4_free_pages(pages, i);
1075 			return NULL;
1076 		}
1077 	}
1078 
1079 	return pages;
1080 }
1081 
1082 static struct nfs4_layoutget *
1083 pnfs_alloc_init_layoutget_args(struct inode *ino,
1084 	   struct nfs_open_context *ctx,
1085 	   const nfs4_stateid *stateid,
1086 	   const struct pnfs_layout_range *range,
1087 	   gfp_t gfp_flags)
1088 {
1089 	struct nfs_server *server = pnfs_find_server(ino, ctx);
1090 	size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1091 	size_t max_pages = max_response_pages(server);
1092 	struct nfs4_layoutget *lgp;
1093 
1094 	dprintk("--> %s\n", __func__);
1095 
1096 	lgp = kzalloc(sizeof(*lgp), gfp_flags);
1097 	if (lgp == NULL)
1098 		return NULL;
1099 
1100 	if (max_reply_sz) {
1101 		size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1102 		if (npages < max_pages)
1103 			max_pages = npages;
1104 	}
1105 
1106 	lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1107 	if (!lgp->args.layout.pages) {
1108 		kfree(lgp);
1109 		return NULL;
1110 	}
1111 	lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1112 	lgp->res.layoutp = &lgp->args.layout;
1113 
1114 	/* Don't confuse uninitialised result and success */
1115 	lgp->res.status = -NFS4ERR_DELAY;
1116 
1117 	lgp->args.minlength = PAGE_SIZE;
1118 	if (lgp->args.minlength > range->length)
1119 		lgp->args.minlength = range->length;
1120 	if (ino) {
1121 		loff_t i_size = i_size_read(ino);
1122 
1123 		if (range->iomode == IOMODE_READ) {
1124 			if (range->offset >= i_size)
1125 				lgp->args.minlength = 0;
1126 			else if (i_size - range->offset < lgp->args.minlength)
1127 				lgp->args.minlength = i_size - range->offset;
1128 		}
1129 	}
1130 	lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1131 	pnfs_copy_range(&lgp->args.range, range);
1132 	lgp->args.type = server->pnfs_curr_ld->id;
1133 	lgp->args.inode = ino;
1134 	lgp->args.ctx = get_nfs_open_context(ctx);
1135 	nfs4_stateid_copy(&lgp->args.stateid, stateid);
1136 	lgp->gfp_flags = gfp_flags;
1137 	lgp->cred = ctx->cred;
1138 	return lgp;
1139 }
1140 
1141 void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1142 {
1143 	size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1144 
1145 	nfs4_free_pages(lgp->args.layout.pages, max_pages);
1146 	pnfs_put_layout_hdr(lgp->lo);
1147 	put_nfs_open_context(lgp->args.ctx);
1148 	kfree(lgp);
1149 }
1150 
1151 static void pnfs_clear_layoutcommit(struct inode *inode,
1152 		struct list_head *head)
1153 {
1154 	struct nfs_inode *nfsi = NFS_I(inode);
1155 	struct pnfs_layout_segment *lseg, *tmp;
1156 
1157 	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1158 		return;
1159 	list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1160 		if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1161 			continue;
1162 		pnfs_lseg_dec_and_remove_zero(lseg, head);
1163 	}
1164 }
1165 
1166 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1167 		const nfs4_stateid *arg_stateid,
1168 		const struct pnfs_layout_range *range,
1169 		const nfs4_stateid *stateid)
1170 {
1171 	struct inode *inode = lo->plh_inode;
1172 	LIST_HEAD(freeme);
1173 
1174 	spin_lock(&inode->i_lock);
1175 	if (!pnfs_layout_is_valid(lo) ||
1176 	    !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1177 		goto out_unlock;
1178 	if (stateid) {
1179 		u32 seq = be32_to_cpu(arg_stateid->seqid);
1180 
1181 		pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1182 		pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1183 		pnfs_set_layout_stateid(lo, stateid, NULL, true);
1184 	} else
1185 		pnfs_mark_layout_stateid_invalid(lo, &freeme);
1186 out_unlock:
1187 	pnfs_clear_layoutreturn_waitbit(lo);
1188 	spin_unlock(&inode->i_lock);
1189 	pnfs_free_lseg_list(&freeme);
1190 
1191 }
1192 
1193 static bool
1194 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1195 		nfs4_stateid *stateid,
1196 		const struct cred **cred,
1197 		enum pnfs_iomode *iomode)
1198 {
1199 	/* Serialise LAYOUTGET/LAYOUTRETURN */
1200 	if (atomic_read(&lo->plh_outstanding) != 0)
1201 		return false;
1202 	if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1203 		return false;
1204 	set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1205 	pnfs_get_layout_hdr(lo);
1206 	nfs4_stateid_copy(stateid, &lo->plh_stateid);
1207 	*cred = get_cred(lo->plh_lc_cred);
1208 	if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1209 		if (lo->plh_return_seq != 0)
1210 			stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1211 		if (iomode != NULL)
1212 			*iomode = lo->plh_return_iomode;
1213 		pnfs_clear_layoutreturn_info(lo);
1214 	} else if (iomode != NULL)
1215 		*iomode = IOMODE_ANY;
1216 	pnfs_barrier_update(lo, be32_to_cpu(stateid->seqid));
1217 	return true;
1218 }
1219 
1220 static void
1221 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1222 		struct pnfs_layout_hdr *lo,
1223 		const nfs4_stateid *stateid,
1224 		enum pnfs_iomode iomode)
1225 {
1226 	struct inode *inode = lo->plh_inode;
1227 
1228 	args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1229 	args->inode = inode;
1230 	args->range.iomode = iomode;
1231 	args->range.offset = 0;
1232 	args->range.length = NFS4_MAX_UINT64;
1233 	args->layout = lo;
1234 	nfs4_stateid_copy(&args->stateid, stateid);
1235 }
1236 
1237 static int
1238 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1239 		       const nfs4_stateid *stateid,
1240 		       const struct cred **pcred,
1241 		       enum pnfs_iomode iomode,
1242 		       bool sync)
1243 {
1244 	struct inode *ino = lo->plh_inode;
1245 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1246 	struct nfs4_layoutreturn *lrp;
1247 	const struct cred *cred = *pcred;
1248 	int status = 0;
1249 
1250 	*pcred = NULL;
1251 	lrp = kzalloc(sizeof(*lrp), nfs_io_gfp_mask());
1252 	if (unlikely(lrp == NULL)) {
1253 		status = -ENOMEM;
1254 		spin_lock(&ino->i_lock);
1255 		pnfs_clear_layoutreturn_waitbit(lo);
1256 		spin_unlock(&ino->i_lock);
1257 		put_cred(cred);
1258 		pnfs_put_layout_hdr(lo);
1259 		goto out;
1260 	}
1261 
1262 	pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1263 	lrp->args.ld_private = &lrp->ld_private;
1264 	lrp->clp = NFS_SERVER(ino)->nfs_client;
1265 	lrp->cred = cred;
1266 	if (ld->prepare_layoutreturn)
1267 		ld->prepare_layoutreturn(&lrp->args);
1268 
1269 	status = nfs4_proc_layoutreturn(lrp, sync);
1270 out:
1271 	dprintk("<-- %s status: %d\n", __func__, status);
1272 	return status;
1273 }
1274 
1275 static bool
1276 pnfs_layout_segments_returnable(struct pnfs_layout_hdr *lo,
1277 				enum pnfs_iomode iomode,
1278 				u32 seq)
1279 {
1280 	struct pnfs_layout_range recall_range = {
1281 		.length = NFS4_MAX_UINT64,
1282 		.iomode = iomode,
1283 	};
1284 	return pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
1285 					       &recall_range, seq) != -EBUSY;
1286 }
1287 
1288 /* Return true if layoutreturn is needed */
1289 static bool
1290 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1291 {
1292 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1293 		return false;
1294 	return pnfs_layout_segments_returnable(lo, lo->plh_return_iomode,
1295 					       lo->plh_return_seq);
1296 }
1297 
1298 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1299 {
1300 	struct inode *inode= lo->plh_inode;
1301 
1302 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1303 		return;
1304 	spin_lock(&inode->i_lock);
1305 	if (pnfs_layout_need_return(lo)) {
1306 		const struct cred *cred;
1307 		nfs4_stateid stateid;
1308 		enum pnfs_iomode iomode;
1309 		bool send;
1310 
1311 		send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1312 		spin_unlock(&inode->i_lock);
1313 		if (send) {
1314 			/* Send an async layoutreturn so we dont deadlock */
1315 			pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
1316 		}
1317 	} else
1318 		spin_unlock(&inode->i_lock);
1319 }
1320 
1321 /*
1322  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1323  * when the layout segment list is empty.
1324  *
1325  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1326  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1327  * deviceid is marked invalid.
1328  */
1329 int
1330 _pnfs_return_layout(struct inode *ino)
1331 {
1332 	struct pnfs_layout_hdr *lo = NULL;
1333 	struct nfs_inode *nfsi = NFS_I(ino);
1334 	struct pnfs_layout_range range = {
1335 		.iomode		= IOMODE_ANY,
1336 		.offset		= 0,
1337 		.length		= NFS4_MAX_UINT64,
1338 	};
1339 	LIST_HEAD(tmp_list);
1340 	const struct cred *cred;
1341 	nfs4_stateid stateid;
1342 	int status = 0;
1343 	bool send, valid_layout;
1344 
1345 	dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1346 
1347 	spin_lock(&ino->i_lock);
1348 	lo = nfsi->layout;
1349 	if (!lo) {
1350 		spin_unlock(&ino->i_lock);
1351 		dprintk("NFS: %s no layout to return\n", __func__);
1352 		goto out;
1353 	}
1354 	/* Reference matched in nfs4_layoutreturn_release */
1355 	pnfs_get_layout_hdr(lo);
1356 	/* Is there an outstanding layoutreturn ? */
1357 	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1358 		spin_unlock(&ino->i_lock);
1359 		if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1360 					TASK_UNINTERRUPTIBLE))
1361 			goto out_put_layout_hdr;
1362 		spin_lock(&ino->i_lock);
1363 	}
1364 	valid_layout = pnfs_layout_is_valid(lo);
1365 	pnfs_clear_layoutcommit(ino, &tmp_list);
1366 	pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1367 
1368 	if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1369 		NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1370 
1371 	/* Don't send a LAYOUTRETURN if list was initially empty */
1372 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1373 			!valid_layout) {
1374 		spin_unlock(&ino->i_lock);
1375 		dprintk("NFS: %s no layout segments to return\n", __func__);
1376 		goto out_wait_layoutreturn;
1377 	}
1378 
1379 	send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1380 	spin_unlock(&ino->i_lock);
1381 	if (send)
1382 		status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY, true);
1383 out_wait_layoutreturn:
1384 	wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1385 out_put_layout_hdr:
1386 	pnfs_free_lseg_list(&tmp_list);
1387 	pnfs_put_layout_hdr(lo);
1388 out:
1389 	dprintk("<-- %s status: %d\n", __func__, status);
1390 	return status;
1391 }
1392 
1393 int
1394 pnfs_commit_and_return_layout(struct inode *inode)
1395 {
1396 	struct pnfs_layout_hdr *lo;
1397 	int ret;
1398 
1399 	spin_lock(&inode->i_lock);
1400 	lo = NFS_I(inode)->layout;
1401 	if (lo == NULL) {
1402 		spin_unlock(&inode->i_lock);
1403 		return 0;
1404 	}
1405 	pnfs_get_layout_hdr(lo);
1406 	/* Block new layoutgets and read/write to ds */
1407 	lo->plh_block_lgets++;
1408 	spin_unlock(&inode->i_lock);
1409 	filemap_fdatawait(inode->i_mapping);
1410 	ret = pnfs_layoutcommit_inode(inode, true);
1411 	if (ret == 0)
1412 		ret = _pnfs_return_layout(inode);
1413 	spin_lock(&inode->i_lock);
1414 	lo->plh_block_lgets--;
1415 	spin_unlock(&inode->i_lock);
1416 	pnfs_put_layout_hdr(lo);
1417 	return ret;
1418 }
1419 
1420 bool pnfs_roc(struct inode *ino,
1421 		struct nfs4_layoutreturn_args *args,
1422 		struct nfs4_layoutreturn_res *res,
1423 		const struct cred *cred)
1424 {
1425 	struct nfs_inode *nfsi = NFS_I(ino);
1426 	struct nfs_open_context *ctx;
1427 	struct nfs4_state *state;
1428 	struct pnfs_layout_hdr *lo;
1429 	struct pnfs_layout_segment *lseg, *next;
1430 	const struct cred *lc_cred;
1431 	nfs4_stateid stateid;
1432 	enum pnfs_iomode iomode = 0;
1433 	bool layoutreturn = false, roc = false;
1434 	bool skip_read = false;
1435 
1436 	if (!nfs_have_layout(ino))
1437 		return false;
1438 retry:
1439 	rcu_read_lock();
1440 	spin_lock(&ino->i_lock);
1441 	lo = nfsi->layout;
1442 	if (!lo || !pnfs_layout_is_valid(lo) ||
1443 	    test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1444 		lo = NULL;
1445 		goto out_noroc;
1446 	}
1447 	pnfs_get_layout_hdr(lo);
1448 	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1449 		spin_unlock(&ino->i_lock);
1450 		rcu_read_unlock();
1451 		wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1452 				TASK_UNINTERRUPTIBLE);
1453 		pnfs_put_layout_hdr(lo);
1454 		goto retry;
1455 	}
1456 
1457 	/* no roc if we hold a delegation */
1458 	if (nfs4_check_delegation(ino, FMODE_READ)) {
1459 		if (nfs4_check_delegation(ino, FMODE_WRITE))
1460 			goto out_noroc;
1461 		skip_read = true;
1462 	}
1463 
1464 	list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1465 		state = ctx->state;
1466 		if (state == NULL)
1467 			continue;
1468 		/* Don't return layout if there is open file state */
1469 		if (state->state & FMODE_WRITE)
1470 			goto out_noroc;
1471 		if (state->state & FMODE_READ)
1472 			skip_read = true;
1473 	}
1474 
1475 
1476 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1477 		if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1478 			continue;
1479 		/* If we are sending layoutreturn, invalidate all valid lsegs */
1480 		if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1481 			continue;
1482 		/*
1483 		 * Note: mark lseg for return so pnfs_layout_remove_lseg
1484 		 * doesn't invalidate the layout for us.
1485 		 */
1486 		set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1487 		if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1488 			continue;
1489 		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1490 	}
1491 
1492 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1493 		goto out_noroc;
1494 
1495 	/* ROC in two conditions:
1496 	 * 1. there are ROC lsegs
1497 	 * 2. we don't send layoutreturn
1498 	 */
1499 	/* lo ref dropped in pnfs_roc_release() */
1500 	layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1501 	/* If the creds don't match, we can't compound the layoutreturn */
1502 	if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1503 		goto out_noroc;
1504 
1505 	roc = layoutreturn;
1506 	pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1507 	res->lrs_present = 0;
1508 	layoutreturn = false;
1509 	put_cred(lc_cred);
1510 
1511 out_noroc:
1512 	spin_unlock(&ino->i_lock);
1513 	rcu_read_unlock();
1514 	pnfs_layoutcommit_inode(ino, true);
1515 	if (roc) {
1516 		struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1517 		if (ld->prepare_layoutreturn)
1518 			ld->prepare_layoutreturn(args);
1519 		pnfs_put_layout_hdr(lo);
1520 		return true;
1521 	}
1522 	if (layoutreturn)
1523 		pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1524 	pnfs_put_layout_hdr(lo);
1525 	return false;
1526 }
1527 
1528 int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1529 		  struct nfs4_layoutreturn_res **respp, int *ret)
1530 {
1531 	struct nfs4_layoutreturn_args *arg = *argpp;
1532 	int retval = -EAGAIN;
1533 
1534 	if (!arg)
1535 		return 0;
1536 	/* Handle Layoutreturn errors */
1537 	switch (*ret) {
1538 	case 0:
1539 		retval = 0;
1540 		break;
1541 	case -NFS4ERR_NOMATCHING_LAYOUT:
1542 		/* Was there an RPC level error? If not, retry */
1543 		if (task->tk_rpc_status == 0)
1544 			break;
1545 		/* If the call was not sent, let caller handle it */
1546 		if (!RPC_WAS_SENT(task))
1547 			return 0;
1548 		/*
1549 		 * Otherwise, assume the call succeeded and
1550 		 * that we need to release the layout
1551 		 */
1552 		*ret = 0;
1553 		(*respp)->lrs_present = 0;
1554 		retval = 0;
1555 		break;
1556 	case -NFS4ERR_DELAY:
1557 		/* Let the caller handle the retry */
1558 		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1559 		return 0;
1560 	case -NFS4ERR_OLD_STATEID:
1561 		if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1562 						     &arg->range, arg->inode))
1563 			break;
1564 		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1565 		return -EAGAIN;
1566 	}
1567 	*argpp = NULL;
1568 	*respp = NULL;
1569 	return retval;
1570 }
1571 
1572 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1573 		struct nfs4_layoutreturn_res *res,
1574 		int ret)
1575 {
1576 	struct pnfs_layout_hdr *lo = args->layout;
1577 	struct inode *inode = args->inode;
1578 	const nfs4_stateid *res_stateid = NULL;
1579 	struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1580 
1581 	switch (ret) {
1582 	case -NFS4ERR_NOMATCHING_LAYOUT:
1583 		spin_lock(&inode->i_lock);
1584 		if (pnfs_layout_is_valid(lo) &&
1585 		    nfs4_stateid_match_other(&args->stateid, &lo->plh_stateid))
1586 			pnfs_set_plh_return_info(lo, args->range.iomode, 0);
1587 		pnfs_clear_layoutreturn_waitbit(lo);
1588 		spin_unlock(&inode->i_lock);
1589 		break;
1590 	case 0:
1591 		if (res->lrs_present)
1592 			res_stateid = &res->stateid;
1593 		fallthrough;
1594 	default:
1595 		pnfs_layoutreturn_free_lsegs(lo, &args->stateid, &args->range,
1596 					     res_stateid);
1597 	}
1598 	trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1599 	if (ld_private && ld_private->ops && ld_private->ops->free)
1600 		ld_private->ops->free(ld_private);
1601 	pnfs_put_layout_hdr(lo);
1602 }
1603 
1604 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1605 {
1606 	struct nfs_inode *nfsi = NFS_I(ino);
1607         struct pnfs_layout_hdr *lo;
1608         bool sleep = false;
1609 
1610 	/* we might not have grabbed lo reference. so need to check under
1611 	 * i_lock */
1612         spin_lock(&ino->i_lock);
1613         lo = nfsi->layout;
1614         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1615                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1616                 sleep = true;
1617 	}
1618         spin_unlock(&ino->i_lock);
1619         return sleep;
1620 }
1621 
1622 /*
1623  * Compare two layout segments for sorting into layout cache.
1624  * We want to preferentially return RW over RO layouts, so ensure those
1625  * are seen first.
1626  */
1627 static s64
1628 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1629 	   const struct pnfs_layout_range *l2)
1630 {
1631 	s64 d;
1632 
1633 	/* high offset > low offset */
1634 	d = l1->offset - l2->offset;
1635 	if (d)
1636 		return d;
1637 
1638 	/* short length > long length */
1639 	d = l2->length - l1->length;
1640 	if (d)
1641 		return d;
1642 
1643 	/* read > read/write */
1644 	return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1645 }
1646 
1647 static bool
1648 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1649 		const struct pnfs_layout_range *l2)
1650 {
1651 	return pnfs_lseg_range_cmp(l1, l2) > 0;
1652 }
1653 
1654 static bool
1655 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1656 		struct pnfs_layout_segment *old)
1657 {
1658 	return false;
1659 }
1660 
1661 void
1662 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1663 		   struct pnfs_layout_segment *lseg,
1664 		   bool (*is_after)(const struct pnfs_layout_range *,
1665 			   const struct pnfs_layout_range *),
1666 		   bool (*do_merge)(struct pnfs_layout_segment *,
1667 			   struct pnfs_layout_segment *),
1668 		   struct list_head *free_me)
1669 {
1670 	struct pnfs_layout_segment *lp, *tmp;
1671 
1672 	dprintk("%s:Begin\n", __func__);
1673 
1674 	list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1675 		if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1676 			continue;
1677 		if (do_merge(lseg, lp)) {
1678 			mark_lseg_invalid(lp, free_me);
1679 			continue;
1680 		}
1681 		if (is_after(&lseg->pls_range, &lp->pls_range))
1682 			continue;
1683 		list_add_tail(&lseg->pls_list, &lp->pls_list);
1684 		dprintk("%s: inserted lseg %p "
1685 			"iomode %d offset %llu length %llu before "
1686 			"lp %p iomode %d offset %llu length %llu\n",
1687 			__func__, lseg, lseg->pls_range.iomode,
1688 			lseg->pls_range.offset, lseg->pls_range.length,
1689 			lp, lp->pls_range.iomode, lp->pls_range.offset,
1690 			lp->pls_range.length);
1691 		goto out;
1692 	}
1693 	list_add_tail(&lseg->pls_list, &lo->plh_segs);
1694 	dprintk("%s: inserted lseg %p "
1695 		"iomode %d offset %llu length %llu at tail\n",
1696 		__func__, lseg, lseg->pls_range.iomode,
1697 		lseg->pls_range.offset, lseg->pls_range.length);
1698 out:
1699 	pnfs_get_layout_hdr(lo);
1700 
1701 	dprintk("%s:Return\n", __func__);
1702 }
1703 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1704 
1705 static void
1706 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1707 		   struct pnfs_layout_segment *lseg,
1708 		   struct list_head *free_me)
1709 {
1710 	struct inode *inode = lo->plh_inode;
1711 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1712 
1713 	if (ld->add_lseg != NULL)
1714 		ld->add_lseg(lo, lseg, free_me);
1715 	else
1716 		pnfs_generic_layout_insert_lseg(lo, lseg,
1717 				pnfs_lseg_range_is_after,
1718 				pnfs_lseg_no_merge,
1719 				free_me);
1720 }
1721 
1722 static struct pnfs_layout_hdr *
1723 alloc_init_layout_hdr(struct inode *ino,
1724 		      struct nfs_open_context *ctx,
1725 		      gfp_t gfp_flags)
1726 {
1727 	struct pnfs_layout_hdr *lo;
1728 
1729 	lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1730 	if (!lo)
1731 		return NULL;
1732 	refcount_set(&lo->plh_refcount, 1);
1733 	INIT_LIST_HEAD(&lo->plh_layouts);
1734 	INIT_LIST_HEAD(&lo->plh_segs);
1735 	INIT_LIST_HEAD(&lo->plh_return_segs);
1736 	INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1737 	lo->plh_inode = ino;
1738 	lo->plh_lc_cred = get_cred(ctx->cred);
1739 	lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1740 	return lo;
1741 }
1742 
1743 static struct pnfs_layout_hdr *
1744 pnfs_find_alloc_layout(struct inode *ino,
1745 		       struct nfs_open_context *ctx,
1746 		       gfp_t gfp_flags)
1747 	__releases(&ino->i_lock)
1748 	__acquires(&ino->i_lock)
1749 {
1750 	struct nfs_inode *nfsi = NFS_I(ino);
1751 	struct pnfs_layout_hdr *new = NULL;
1752 
1753 	dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1754 
1755 	if (nfsi->layout != NULL)
1756 		goto out_existing;
1757 	spin_unlock(&ino->i_lock);
1758 	new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1759 	spin_lock(&ino->i_lock);
1760 
1761 	if (likely(nfsi->layout == NULL)) {	/* Won the race? */
1762 		nfsi->layout = new;
1763 		return new;
1764 	} else if (new != NULL)
1765 		pnfs_free_layout_hdr(new);
1766 out_existing:
1767 	pnfs_get_layout_hdr(nfsi->layout);
1768 	return nfsi->layout;
1769 }
1770 
1771 /*
1772  * iomode matching rules:
1773  * iomode	lseg	strict match
1774  *                      iomode
1775  * -----	-----	------ -----
1776  * ANY		READ	N/A    true
1777  * ANY		RW	N/A    true
1778  * RW		READ	N/A    false
1779  * RW		RW	N/A    true
1780  * READ		READ	N/A    true
1781  * READ		RW	true   false
1782  * READ		RW	false  true
1783  */
1784 static bool
1785 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1786 		 const struct pnfs_layout_range *range,
1787 		 bool strict_iomode)
1788 {
1789 	struct pnfs_layout_range range1;
1790 
1791 	if ((range->iomode == IOMODE_RW &&
1792 	     ls_range->iomode != IOMODE_RW) ||
1793 	    (range->iomode != ls_range->iomode &&
1794 	     strict_iomode) ||
1795 	    !pnfs_lseg_range_intersecting(ls_range, range))
1796 		return false;
1797 
1798 	/* range1 covers only the first byte in the range */
1799 	range1 = *range;
1800 	range1.length = 1;
1801 	return pnfs_lseg_range_contained(ls_range, &range1);
1802 }
1803 
1804 /*
1805  * lookup range in layout
1806  */
1807 static struct pnfs_layout_segment *
1808 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1809 		struct pnfs_layout_range *range,
1810 		bool strict_iomode)
1811 {
1812 	struct pnfs_layout_segment *lseg, *ret = NULL;
1813 
1814 	dprintk("%s:Begin\n", __func__);
1815 
1816 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1817 		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1818 		    pnfs_lseg_range_match(&lseg->pls_range, range,
1819 					  strict_iomode)) {
1820 			ret = pnfs_get_lseg(lseg);
1821 			break;
1822 		}
1823 	}
1824 
1825 	dprintk("%s:Return lseg %p ref %d\n",
1826 		__func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1827 	return ret;
1828 }
1829 
1830 /*
1831  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1832  * to the MDS or over pNFS
1833  *
1834  * The nfs_inode read_io and write_io fields are cumulative counters reset
1835  * when there are no layout segments. Note that in pnfs_update_layout iomode
1836  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1837  * WRITE request.
1838  *
1839  * A return of true means use MDS I/O.
1840  *
1841  * From rfc 5661:
1842  * If a file's size is smaller than the file size threshold, data accesses
1843  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1844  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1845  * server.  If both file size and I/O size are provided, the client SHOULD
1846  * reach or exceed  both thresholds before sending its read or write
1847  * requests to the data server.
1848  */
1849 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1850 				     struct inode *ino, int iomode)
1851 {
1852 	struct nfs4_threshold *t = ctx->mdsthreshold;
1853 	struct nfs_inode *nfsi = NFS_I(ino);
1854 	loff_t fsize = i_size_read(ino);
1855 	bool size = false, size_set = false, io = false, io_set = false, ret = false;
1856 
1857 	if (t == NULL)
1858 		return ret;
1859 
1860 	dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1861 		__func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1862 
1863 	switch (iomode) {
1864 	case IOMODE_READ:
1865 		if (t->bm & THRESHOLD_RD) {
1866 			dprintk("%s fsize %llu\n", __func__, fsize);
1867 			size_set = true;
1868 			if (fsize < t->rd_sz)
1869 				size = true;
1870 		}
1871 		if (t->bm & THRESHOLD_RD_IO) {
1872 			dprintk("%s nfsi->read_io %llu\n", __func__,
1873 				nfsi->read_io);
1874 			io_set = true;
1875 			if (nfsi->read_io < t->rd_io_sz)
1876 				io = true;
1877 		}
1878 		break;
1879 	case IOMODE_RW:
1880 		if (t->bm & THRESHOLD_WR) {
1881 			dprintk("%s fsize %llu\n", __func__, fsize);
1882 			size_set = true;
1883 			if (fsize < t->wr_sz)
1884 				size = true;
1885 		}
1886 		if (t->bm & THRESHOLD_WR_IO) {
1887 			dprintk("%s nfsi->write_io %llu\n", __func__,
1888 				nfsi->write_io);
1889 			io_set = true;
1890 			if (nfsi->write_io < t->wr_io_sz)
1891 				io = true;
1892 		}
1893 		break;
1894 	}
1895 	if (size_set && io_set) {
1896 		if (size && io)
1897 			ret = true;
1898 	} else if (size || io)
1899 		ret = true;
1900 
1901 	dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1902 	return ret;
1903 }
1904 
1905 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1906 {
1907 	/*
1908 	 * send layoutcommit as it can hold up layoutreturn due to lseg
1909 	 * reference
1910 	 */
1911 	pnfs_layoutcommit_inode(lo->plh_inode, false);
1912 	return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1913 				   nfs_wait_bit_killable,
1914 				   TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
1915 }
1916 
1917 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1918 {
1919 	atomic_inc(&lo->plh_outstanding);
1920 }
1921 
1922 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1923 {
1924 	if (atomic_dec_and_test(&lo->plh_outstanding) &&
1925 	    test_and_clear_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags))
1926 		wake_up_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN);
1927 }
1928 
1929 static bool pnfs_is_first_layoutget(struct pnfs_layout_hdr *lo)
1930 {
1931 	return test_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags);
1932 }
1933 
1934 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1935 {
1936 	unsigned long *bitlock = &lo->plh_flags;
1937 
1938 	clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1939 	smp_mb__after_atomic();
1940 	wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1941 }
1942 
1943 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1944 				struct nfs_server *server)
1945 {
1946 	if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1947 		struct nfs_client *clp = server->nfs_client;
1948 
1949 		/* The lo must be on the clp list if there is any
1950 		 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1951 		 */
1952 		spin_lock(&clp->cl_lock);
1953 		list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1954 		spin_unlock(&clp->cl_lock);
1955 	}
1956 }
1957 
1958 /*
1959  * Layout segment is retreived from the server if not cached.
1960  * The appropriate layout segment is referenced and returned to the caller.
1961  */
1962 struct pnfs_layout_segment *
1963 pnfs_update_layout(struct inode *ino,
1964 		   struct nfs_open_context *ctx,
1965 		   loff_t pos,
1966 		   u64 count,
1967 		   enum pnfs_iomode iomode,
1968 		   bool strict_iomode,
1969 		   gfp_t gfp_flags)
1970 {
1971 	struct pnfs_layout_range arg = {
1972 		.iomode = iomode,
1973 		.offset = pos,
1974 		.length = count,
1975 	};
1976 	unsigned pg_offset;
1977 	struct nfs_server *server = NFS_SERVER(ino);
1978 	struct nfs_client *clp = server->nfs_client;
1979 	struct pnfs_layout_hdr *lo = NULL;
1980 	struct pnfs_layout_segment *lseg = NULL;
1981 	struct nfs4_layoutget *lgp;
1982 	nfs4_stateid stateid;
1983 	struct nfs4_exception exception = {
1984 		.inode = ino,
1985 	};
1986 	unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1987 	bool first;
1988 
1989 	if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1990 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1991 				 PNFS_UPDATE_LAYOUT_NO_PNFS);
1992 		goto out;
1993 	}
1994 
1995 	if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1996 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1997 				 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1998 		goto out;
1999 	}
2000 
2001 lookup_again:
2002 	lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
2003 	if (IS_ERR(lseg))
2004 		goto out;
2005 	first = false;
2006 	spin_lock(&ino->i_lock);
2007 	lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
2008 	if (lo == NULL) {
2009 		spin_unlock(&ino->i_lock);
2010 		lseg = ERR_PTR(-ENOMEM);
2011 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2012 				 PNFS_UPDATE_LAYOUT_NOMEM);
2013 		goto out;
2014 	}
2015 
2016 	/* Do we even need to bother with this? */
2017 	if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
2018 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2019 				 PNFS_UPDATE_LAYOUT_BULK_RECALL);
2020 		dprintk("%s matches recall, use MDS\n", __func__);
2021 		goto out_unlock;
2022 	}
2023 
2024 	/* if LAYOUTGET already failed once we don't try again */
2025 	if (pnfs_layout_io_test_failed(lo, iomode)) {
2026 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2027 				 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2028 		goto out_unlock;
2029 	}
2030 
2031 	/*
2032 	 * If the layout segment list is empty, but there are outstanding
2033 	 * layoutget calls, then they might be subject to a layoutrecall.
2034 	 */
2035 	if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2036 	    atomic_read(&lo->plh_outstanding) != 0) {
2037 		spin_unlock(&ino->i_lock);
2038 		lseg = ERR_PTR(wait_on_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN,
2039 					   TASK_KILLABLE));
2040 		if (IS_ERR(lseg))
2041 			goto out_put_layout_hdr;
2042 		pnfs_put_layout_hdr(lo);
2043 		goto lookup_again;
2044 	}
2045 
2046 	/*
2047 	 * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2048 	 * for LAYOUTRETURN.
2049 	 */
2050 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2051 		spin_unlock(&ino->i_lock);
2052 		dprintk("%s wait for layoutreturn\n", __func__);
2053 		lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2054 		if (!IS_ERR(lseg)) {
2055 			pnfs_put_layout_hdr(lo);
2056 			dprintk("%s retrying\n", __func__);
2057 			trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2058 						 lseg,
2059 						 PNFS_UPDATE_LAYOUT_RETRY);
2060 			goto lookup_again;
2061 		}
2062 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2063 					 PNFS_UPDATE_LAYOUT_RETURN);
2064 		goto out_put_layout_hdr;
2065 	}
2066 
2067 	lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2068 	if (lseg) {
2069 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2070 				PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2071 		goto out_unlock;
2072 	}
2073 
2074 	/*
2075 	 * Choose a stateid for the LAYOUTGET. If we don't have a layout
2076 	 * stateid, or it has been invalidated, then we must use the open
2077 	 * stateid.
2078 	 */
2079 	if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2080 		int status;
2081 
2082 		/*
2083 		 * The first layoutget for the file. Need to serialize per
2084 		 * RFC 5661 Errata 3208.
2085 		 */
2086 		if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2087 				     &lo->plh_flags)) {
2088 			spin_unlock(&ino->i_lock);
2089 			lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2090 						NFS_LAYOUT_FIRST_LAYOUTGET,
2091 						TASK_KILLABLE));
2092 			if (IS_ERR(lseg))
2093 				goto out_put_layout_hdr;
2094 			pnfs_put_layout_hdr(lo);
2095 			dprintk("%s retrying\n", __func__);
2096 			goto lookup_again;
2097 		}
2098 
2099 		spin_unlock(&ino->i_lock);
2100 		first = true;
2101 		status = nfs4_select_rw_stateid(ctx->state,
2102 					iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2103 					NULL, &stateid, NULL);
2104 		if (status != 0) {
2105 			lseg = ERR_PTR(status);
2106 			trace_pnfs_update_layout(ino, pos, count,
2107 					iomode, lo, lseg,
2108 					PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2109 			nfs4_schedule_stateid_recovery(server, ctx->state);
2110 			pnfs_clear_first_layoutget(lo);
2111 			pnfs_put_layout_hdr(lo);
2112 			goto lookup_again;
2113 		}
2114 		spin_lock(&ino->i_lock);
2115 	} else {
2116 		nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2117 	}
2118 
2119 	if (pnfs_layoutgets_blocked(lo)) {
2120 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2121 				PNFS_UPDATE_LAYOUT_BLOCKED);
2122 		goto out_unlock;
2123 	}
2124 	nfs_layoutget_begin(lo);
2125 	spin_unlock(&ino->i_lock);
2126 
2127 	_add_to_server_list(lo, server);
2128 
2129 	pg_offset = arg.offset & ~PAGE_MASK;
2130 	if (pg_offset) {
2131 		arg.offset -= pg_offset;
2132 		arg.length += pg_offset;
2133 	}
2134 	if (arg.length != NFS4_MAX_UINT64)
2135 		arg.length = PAGE_ALIGN(arg.length);
2136 
2137 	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2138 	if (!lgp) {
2139 		lseg = ERR_PTR(-ENOMEM);
2140 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2141 					 PNFS_UPDATE_LAYOUT_NOMEM);
2142 		nfs_layoutget_end(lo);
2143 		goto out_put_layout_hdr;
2144 	}
2145 
2146 	lgp->lo = lo;
2147 	pnfs_get_layout_hdr(lo);
2148 
2149 	lseg = nfs4_proc_layoutget(lgp, &exception);
2150 	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2151 				 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2152 	nfs_layoutget_end(lo);
2153 	if (IS_ERR(lseg)) {
2154 		switch(PTR_ERR(lseg)) {
2155 		case -EBUSY:
2156 			if (time_after(jiffies, giveup))
2157 				lseg = NULL;
2158 			break;
2159 		case -ERECALLCONFLICT:
2160 		case -EAGAIN:
2161 			break;
2162 		case -ENODATA:
2163 			/* The server returned NFS4ERR_LAYOUTUNAVAILABLE */
2164 			pnfs_layout_set_fail_bit(
2165 				lo, pnfs_iomode_to_fail_bit(iomode));
2166 			lseg = NULL;
2167 			goto out_put_layout_hdr;
2168 		default:
2169 			if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2170 				pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2171 				lseg = NULL;
2172 			}
2173 			goto out_put_layout_hdr;
2174 		}
2175 		if (lseg) {
2176 			if (!exception.retry)
2177 				goto out_put_layout_hdr;
2178 			if (first)
2179 				pnfs_clear_first_layoutget(lo);
2180 			trace_pnfs_update_layout(ino, pos, count,
2181 				iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2182 			pnfs_put_layout_hdr(lo);
2183 			goto lookup_again;
2184 		}
2185 	} else {
2186 		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2187 	}
2188 
2189 out_put_layout_hdr:
2190 	if (first)
2191 		pnfs_clear_first_layoutget(lo);
2192 	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2193 				 PNFS_UPDATE_LAYOUT_EXIT);
2194 	pnfs_put_layout_hdr(lo);
2195 out:
2196 	dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2197 			"(%s, offset: %llu, length: %llu)\n",
2198 			__func__, ino->i_sb->s_id,
2199 			(unsigned long long)NFS_FILEID(ino),
2200 			IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2201 			iomode==IOMODE_RW ?  "read/write" : "read-only",
2202 			(unsigned long long)pos,
2203 			(unsigned long long)count);
2204 	return lseg;
2205 out_unlock:
2206 	spin_unlock(&ino->i_lock);
2207 	goto out_put_layout_hdr;
2208 }
2209 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2210 
2211 static bool
2212 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2213 {
2214 	switch (range->iomode) {
2215 	case IOMODE_READ:
2216 	case IOMODE_RW:
2217 		break;
2218 	default:
2219 		return false;
2220 	}
2221 	if (range->offset == NFS4_MAX_UINT64)
2222 		return false;
2223 	if (range->length == 0)
2224 		return false;
2225 	if (range->length != NFS4_MAX_UINT64 &&
2226 	    range->length > NFS4_MAX_UINT64 - range->offset)
2227 		return false;
2228 	return true;
2229 }
2230 
2231 static struct pnfs_layout_hdr *
2232 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2233 {
2234 	struct pnfs_layout_hdr *lo;
2235 
2236 	spin_lock(&ino->i_lock);
2237 	lo = pnfs_find_alloc_layout(ino, ctx, nfs_io_gfp_mask());
2238 	if (!lo)
2239 		goto out_unlock;
2240 	if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2241 		goto out_unlock;
2242 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2243 		goto out_unlock;
2244 	if (pnfs_layoutgets_blocked(lo))
2245 		goto out_unlock;
2246 	if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2247 		goto out_unlock;
2248 	nfs_layoutget_begin(lo);
2249 	spin_unlock(&ino->i_lock);
2250 	_add_to_server_list(lo, NFS_SERVER(ino));
2251 	return lo;
2252 
2253 out_unlock:
2254 	spin_unlock(&ino->i_lock);
2255 	pnfs_put_layout_hdr(lo);
2256 	return NULL;
2257 }
2258 
2259 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2260 				     struct nfs_open_context *ctx)
2261 {
2262 	struct inode *ino = data->dentry->d_inode;
2263 	struct pnfs_layout_range rng = {
2264 		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2265 			  IOMODE_RW: IOMODE_READ,
2266 		.offset = 0,
2267 		.length = NFS4_MAX_UINT64,
2268 	};
2269 	struct nfs4_layoutget *lgp;
2270 	struct pnfs_layout_hdr *lo;
2271 
2272 	/* Heuristic: don't send layoutget if we have cached data */
2273 	if (rng.iomode == IOMODE_READ &&
2274 	   (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2275 		return;
2276 
2277 	lo = _pnfs_grab_empty_layout(ino, ctx);
2278 	if (!lo)
2279 		return;
2280 	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2281 					     nfs_io_gfp_mask());
2282 	if (!lgp) {
2283 		pnfs_clear_first_layoutget(lo);
2284 		nfs_layoutget_end(lo);
2285 		pnfs_put_layout_hdr(lo);
2286 		return;
2287 	}
2288 	lgp->lo = lo;
2289 	data->lgp = lgp;
2290 	data->o_arg.lg_args = &lgp->args;
2291 	data->o_res.lg_res = &lgp->res;
2292 }
2293 
2294 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2295 				     struct nfs_open_context *ctx)
2296 {
2297 	struct inode *ino = data->dentry->d_inode;
2298 	struct pnfs_layout_range rng = {
2299 		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2300 			  IOMODE_RW: IOMODE_READ,
2301 		.offset = 0,
2302 		.length = NFS4_MAX_UINT64,
2303 	};
2304 	struct nfs4_layoutget *lgp;
2305 
2306 	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2307 					     nfs_io_gfp_mask());
2308 	if (!lgp)
2309 		return;
2310 	data->lgp = lgp;
2311 	data->o_arg.lg_args = &lgp->args;
2312 	data->o_res.lg_res = &lgp->res;
2313 }
2314 
2315 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2316 			 struct nfs_open_context *ctx)
2317 {
2318 	struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2319 
2320 	if (!(pnfs_enabled_sb(server) &&
2321 	      server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2322 		return;
2323 	/* Could check on max_ops, but currently hardcoded high enough */
2324 	if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2325 		return;
2326 	if (data->lgp)
2327 		return;
2328 	if (data->state)
2329 		_lgopen_prepare_attached(data, ctx);
2330 	else
2331 		_lgopen_prepare_floating(data, ctx);
2332 }
2333 
2334 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2335 		       struct nfs_open_context *ctx)
2336 {
2337 	struct pnfs_layout_hdr *lo;
2338 	struct pnfs_layout_segment *lseg;
2339 	struct nfs_server *srv = NFS_SERVER(ino);
2340 	u32 iomode;
2341 
2342 	if (!lgp)
2343 		return;
2344 	dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2345 	if (lgp->res.status) {
2346 		switch (lgp->res.status) {
2347 		default:
2348 			break;
2349 		/*
2350 		 * Halt lgopen attempts if the server doesn't recognise
2351 		 * the "current stateid" value, the layout type, or the
2352 		 * layoutget operation as being valid.
2353 		 * Also if it complains about too many ops in the compound
2354 		 * or of the request/reply being too big.
2355 		 */
2356 		case -NFS4ERR_BAD_STATEID:
2357 		case -NFS4ERR_NOTSUPP:
2358 		case -NFS4ERR_REP_TOO_BIG:
2359 		case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2360 		case -NFS4ERR_REQ_TOO_BIG:
2361 		case -NFS4ERR_TOO_MANY_OPS:
2362 		case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2363 			srv->caps &= ~NFS_CAP_LGOPEN;
2364 		}
2365 		return;
2366 	}
2367 	if (!lgp->lo) {
2368 		lo = _pnfs_grab_empty_layout(ino, ctx);
2369 		if (!lo)
2370 			return;
2371 		lgp->lo = lo;
2372 	} else
2373 		lo = lgp->lo;
2374 
2375 	lseg = pnfs_layout_process(lgp);
2376 	if (!IS_ERR(lseg)) {
2377 		iomode = lgp->args.range.iomode;
2378 		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2379 		pnfs_put_lseg(lseg);
2380 	}
2381 }
2382 
2383 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2384 {
2385 	if (lgp != NULL) {
2386 		if (lgp->lo) {
2387 			pnfs_clear_first_layoutget(lgp->lo);
2388 			nfs_layoutget_end(lgp->lo);
2389 		}
2390 		pnfs_layoutget_free(lgp);
2391 	}
2392 }
2393 
2394 struct pnfs_layout_segment *
2395 pnfs_layout_process(struct nfs4_layoutget *lgp)
2396 {
2397 	struct pnfs_layout_hdr *lo = lgp->lo;
2398 	struct nfs4_layoutget_res *res = &lgp->res;
2399 	struct pnfs_layout_segment *lseg;
2400 	struct inode *ino = lo->plh_inode;
2401 	LIST_HEAD(free_me);
2402 
2403 	if (!pnfs_sanity_check_layout_range(&res->range))
2404 		return ERR_PTR(-EINVAL);
2405 
2406 	/* Inject layout blob into I/O device driver */
2407 	lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2408 	if (IS_ERR_OR_NULL(lseg)) {
2409 		if (!lseg)
2410 			lseg = ERR_PTR(-ENOMEM);
2411 
2412 		dprintk("%s: Could not allocate layout: error %ld\n",
2413 		       __func__, PTR_ERR(lseg));
2414 		return lseg;
2415 	}
2416 
2417 	pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2418 
2419 	spin_lock(&ino->i_lock);
2420 	if (pnfs_layoutgets_blocked(lo)) {
2421 		dprintk("%s forget reply due to state\n", __func__);
2422 		goto out_forget;
2423 	}
2424 
2425 	if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2426 	    !pnfs_is_first_layoutget(lo))
2427 		goto out_forget;
2428 
2429 	if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2430 		/* existing state ID, make sure the sequence number matches. */
2431 		if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2432 			if (!pnfs_layout_is_valid(lo))
2433 				lo->plh_barrier = 0;
2434 			dprintk("%s forget reply due to sequence\n", __func__);
2435 			goto out_forget;
2436 		}
2437 		pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2438 	} else if (pnfs_layout_is_valid(lo)) {
2439 		/*
2440 		 * We got an entirely new state ID.  Mark all segments for the
2441 		 * inode invalid, and retry the layoutget
2442 		 */
2443 		struct pnfs_layout_range range = {
2444 			.iomode = IOMODE_ANY,
2445 			.length = NFS4_MAX_UINT64,
2446 		};
2447 		pnfs_mark_matching_lsegs_return(lo, &free_me, &range, 0);
2448 		goto out_forget;
2449 	} else {
2450 		/* We have a completely new layout */
2451 		pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2452 	}
2453 
2454 	pnfs_get_lseg(lseg);
2455 	pnfs_layout_insert_lseg(lo, lseg, &free_me);
2456 
2457 
2458 	if (res->return_on_close)
2459 		set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2460 
2461 	spin_unlock(&ino->i_lock);
2462 	pnfs_free_lseg_list(&free_me);
2463 	return lseg;
2464 
2465 out_forget:
2466 	spin_unlock(&ino->i_lock);
2467 	lseg->pls_layout = lo;
2468 	NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2469 	return ERR_PTR(-EAGAIN);
2470 }
2471 
2472 /**
2473  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2474  * @lo: pointer to layout header
2475  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2476  * @return_range: describe layout segment ranges to be returned
2477  * @seq: stateid seqid to match
2478  *
2479  * This function is mainly intended for use by layoutrecall. It attempts
2480  * to free the layout segment immediately, or else to mark it for return
2481  * as soon as its reference count drops to zero.
2482  *
2483  * Returns
2484  * - 0: a layoutreturn needs to be scheduled.
2485  * - EBUSY: there are layout segment that are still in use.
2486  * - ENOENT: there are no layout segments that need to be returned.
2487  */
2488 int
2489 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2490 				struct list_head *tmp_list,
2491 				const struct pnfs_layout_range *return_range,
2492 				u32 seq)
2493 {
2494 	struct pnfs_layout_segment *lseg, *next;
2495 	struct nfs_server *server = NFS_SERVER(lo->plh_inode);
2496 	int remaining = 0;
2497 
2498 	dprintk("%s:Begin lo %p\n", __func__, lo);
2499 
2500 	assert_spin_locked(&lo->plh_inode->i_lock);
2501 
2502 	if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2503 		tmp_list = &lo->plh_return_segs;
2504 
2505 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2506 		if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2507 			dprintk("%s: marking lseg %p iomode %d "
2508 				"offset %llu length %llu\n", __func__,
2509 				lseg, lseg->pls_range.iomode,
2510 				lseg->pls_range.offset,
2511 				lseg->pls_range.length);
2512 			if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2513 				tmp_list = &lo->plh_return_segs;
2514 			if (mark_lseg_invalid(lseg, tmp_list))
2515 				continue;
2516 			remaining++;
2517 			set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2518 			pnfs_lseg_cancel_io(server, lseg);
2519 		}
2520 
2521 	if (remaining) {
2522 		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2523 		return -EBUSY;
2524 	}
2525 
2526 	if (!list_empty(&lo->plh_return_segs)) {
2527 		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2528 		return 0;
2529 	}
2530 
2531 	return -ENOENT;
2532 }
2533 
2534 static void
2535 pnfs_mark_layout_for_return(struct inode *inode,
2536 			    const struct pnfs_layout_range *range)
2537 {
2538 	struct pnfs_layout_hdr *lo;
2539 	bool return_now = false;
2540 
2541 	spin_lock(&inode->i_lock);
2542 	lo = NFS_I(inode)->layout;
2543 	if (!pnfs_layout_is_valid(lo)) {
2544 		spin_unlock(&inode->i_lock);
2545 		return;
2546 	}
2547 	pnfs_set_plh_return_info(lo, range->iomode, 0);
2548 	/*
2549 	 * mark all matching lsegs so that we are sure to have no live
2550 	 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2551 	 * for how it works.
2552 	 */
2553 	if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2554 		const struct cred *cred;
2555 		nfs4_stateid stateid;
2556 		enum pnfs_iomode iomode;
2557 
2558 		return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2559 		spin_unlock(&inode->i_lock);
2560 		if (return_now)
2561 			pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2562 	} else {
2563 		spin_unlock(&inode->i_lock);
2564 		nfs_commit_inode(inode, 0);
2565 	}
2566 }
2567 
2568 void pnfs_error_mark_layout_for_return(struct inode *inode,
2569 				       struct pnfs_layout_segment *lseg)
2570 {
2571 	struct pnfs_layout_range range = {
2572 		.iomode = lseg->pls_range.iomode,
2573 		.offset = 0,
2574 		.length = NFS4_MAX_UINT64,
2575 	};
2576 
2577 	pnfs_mark_layout_for_return(inode, &range);
2578 }
2579 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2580 
2581 static bool
2582 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2583 {
2584 	return pnfs_layout_is_valid(lo) &&
2585 		!test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2586 		!test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2587 }
2588 
2589 static struct pnfs_layout_segment *
2590 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2591 		     const struct pnfs_layout_range *range,
2592 		     enum pnfs_iomode iomode)
2593 {
2594 	struct pnfs_layout_segment *lseg;
2595 
2596 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2597 		if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2598 			continue;
2599 		if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2600 			continue;
2601 		if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2602 			continue;
2603 		if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2604 			return lseg;
2605 	}
2606 	return NULL;
2607 }
2608 
2609 /* Find open file states whose mode matches that of the range */
2610 static bool
2611 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2612 				 const struct pnfs_layout_range *range)
2613 {
2614 	struct list_head *head;
2615 	struct nfs_open_context *ctx;
2616 	fmode_t mode = 0;
2617 
2618 	if (!pnfs_layout_can_be_returned(lo) ||
2619 	    !pnfs_find_first_lseg(lo, range, range->iomode))
2620 		return false;
2621 
2622 	head = &NFS_I(lo->plh_inode)->open_files;
2623 	list_for_each_entry_rcu(ctx, head, list) {
2624 		if (ctx->state)
2625 			mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2626 	}
2627 
2628 	switch (range->iomode) {
2629 	default:
2630 		break;
2631 	case IOMODE_READ:
2632 		mode &= ~FMODE_WRITE;
2633 		break;
2634 	case IOMODE_RW:
2635 		if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2636 			mode &= ~FMODE_READ;
2637 	}
2638 	return mode == 0;
2639 }
2640 
2641 static int pnfs_layout_return_unused_byserver(struct nfs_server *server,
2642 					      void *data)
2643 {
2644 	const struct pnfs_layout_range *range = data;
2645 	const struct cred *cred;
2646 	struct pnfs_layout_hdr *lo;
2647 	struct inode *inode;
2648 	nfs4_stateid stateid;
2649 	enum pnfs_iomode iomode;
2650 
2651 restart:
2652 	rcu_read_lock();
2653 	list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2654 		inode = lo->plh_inode;
2655 		if (!inode || !pnfs_layout_can_be_returned(lo) ||
2656 		    test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2657 			continue;
2658 		spin_lock(&inode->i_lock);
2659 		if (!lo->plh_inode ||
2660 		    !pnfs_should_return_unused_layout(lo, range)) {
2661 			spin_unlock(&inode->i_lock);
2662 			continue;
2663 		}
2664 		pnfs_get_layout_hdr(lo);
2665 		pnfs_set_plh_return_info(lo, range->iomode, 0);
2666 		if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
2667 						    range, 0) != 0 ||
2668 		    !pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode)) {
2669 			spin_unlock(&inode->i_lock);
2670 			rcu_read_unlock();
2671 			pnfs_put_layout_hdr(lo);
2672 			cond_resched();
2673 			goto restart;
2674 		}
2675 		spin_unlock(&inode->i_lock);
2676 		rcu_read_unlock();
2677 		pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2678 		pnfs_put_layout_hdr(lo);
2679 		cond_resched();
2680 		goto restart;
2681 	}
2682 	rcu_read_unlock();
2683 	return 0;
2684 }
2685 
2686 void
2687 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2688 				 enum pnfs_iomode iomode)
2689 {
2690 	struct pnfs_layout_range range = {
2691 		.iomode = iomode,
2692 		.offset = 0,
2693 		.length = NFS4_MAX_UINT64,
2694 	};
2695 
2696 	nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2697 			&range);
2698 }
2699 
2700 void
2701 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2702 {
2703 	if (pgio->pg_lseg == NULL ||
2704 	    test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2705 		return;
2706 	pnfs_put_lseg(pgio->pg_lseg);
2707 	pgio->pg_lseg = NULL;
2708 }
2709 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2710 
2711 /*
2712  * Check for any intersection between the request and the pgio->pg_lseg,
2713  * and if none, put this pgio->pg_lseg away.
2714  */
2715 void
2716 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2717 {
2718 	if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2719 		pnfs_put_lseg(pgio->pg_lseg);
2720 		pgio->pg_lseg = NULL;
2721 	}
2722 }
2723 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2724 
2725 void
2726 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2727 {
2728 	u64 rd_size;
2729 
2730 	pnfs_generic_pg_check_layout(pgio);
2731 	pnfs_generic_pg_check_range(pgio, req);
2732 	if (pgio->pg_lseg == NULL) {
2733 		if (pgio->pg_dreq == NULL)
2734 			rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2735 		else
2736 			rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2737 
2738 		pgio->pg_lseg =
2739 			pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2740 					   req_offset(req), rd_size,
2741 					   IOMODE_READ, false,
2742 					   nfs_io_gfp_mask());
2743 		if (IS_ERR(pgio->pg_lseg)) {
2744 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2745 			pgio->pg_lseg = NULL;
2746 			return;
2747 		}
2748 	}
2749 	/* If no lseg, fall back to read through mds */
2750 	if (pgio->pg_lseg == NULL)
2751 		nfs_pageio_reset_read_mds(pgio);
2752 
2753 }
2754 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2755 
2756 void
2757 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2758 			   struct nfs_page *req, u64 wb_size)
2759 {
2760 	pnfs_generic_pg_check_layout(pgio);
2761 	pnfs_generic_pg_check_range(pgio, req);
2762 	if (pgio->pg_lseg == NULL) {
2763 		pgio->pg_lseg =
2764 			pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2765 					   req_offset(req), wb_size, IOMODE_RW,
2766 					   false, nfs_io_gfp_mask());
2767 		if (IS_ERR(pgio->pg_lseg)) {
2768 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2769 			pgio->pg_lseg = NULL;
2770 			return;
2771 		}
2772 	}
2773 	/* If no lseg, fall back to write through mds */
2774 	if (pgio->pg_lseg == NULL)
2775 		nfs_pageio_reset_write_mds(pgio);
2776 }
2777 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2778 
2779 void
2780 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2781 {
2782 	if (desc->pg_lseg) {
2783 		pnfs_put_lseg(desc->pg_lseg);
2784 		desc->pg_lseg = NULL;
2785 	}
2786 }
2787 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2788 
2789 /*
2790  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2791  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2792  */
2793 size_t
2794 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2795 		     struct nfs_page *prev, struct nfs_page *req)
2796 {
2797 	unsigned int size;
2798 	u64 seg_end, req_start, seg_left;
2799 
2800 	size = nfs_generic_pg_test(pgio, prev, req);
2801 	if (!size)
2802 		return 0;
2803 
2804 	/*
2805 	 * 'size' contains the number of bytes left in the current page (up
2806 	 * to the original size asked for in @req->wb_bytes).
2807 	 *
2808 	 * Calculate how many bytes are left in the layout segment
2809 	 * and if there are less bytes than 'size', return that instead.
2810 	 *
2811 	 * Please also note that 'end_offset' is actually the offset of the
2812 	 * first byte that lies outside the pnfs_layout_range. FIXME?
2813 	 *
2814 	 */
2815 	if (pgio->pg_lseg) {
2816 		seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2817 				     pgio->pg_lseg->pls_range.length);
2818 		req_start = req_offset(req);
2819 
2820 		/* start of request is past the last byte of this segment */
2821 		if (req_start >= seg_end)
2822 			return 0;
2823 
2824 		/* adjust 'size' iff there are fewer bytes left in the
2825 		 * segment than what nfs_generic_pg_test returned */
2826 		seg_left = seg_end - req_start;
2827 		if (seg_left < size)
2828 			size = (unsigned int)seg_left;
2829 	}
2830 
2831 	return size;
2832 }
2833 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2834 
2835 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2836 {
2837 	struct nfs_pageio_descriptor pgio;
2838 
2839 	/* Resend all requests through the MDS */
2840 	nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2841 			      hdr->completion_ops);
2842 	return nfs_pageio_resend(&pgio, hdr);
2843 }
2844 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2845 
2846 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2847 {
2848 
2849 	dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2850 	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2851 	    PNFS_LAYOUTRET_ON_ERROR) {
2852 		pnfs_return_layout(hdr->inode);
2853 	}
2854 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2855 		hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2856 }
2857 
2858 /*
2859  * Called by non rpc-based layout drivers
2860  */
2861 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2862 {
2863 	if (likely(!hdr->pnfs_error)) {
2864 		pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2865 				hdr->mds_offset + hdr->res.count);
2866 		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2867 	}
2868 	trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2869 	if (unlikely(hdr->pnfs_error))
2870 		pnfs_ld_handle_write_error(hdr);
2871 	hdr->mds_ops->rpc_release(hdr);
2872 }
2873 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2874 
2875 static void
2876 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2877 		struct nfs_pgio_header *hdr)
2878 {
2879 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2880 
2881 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2882 		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2883 		nfs_pageio_reset_write_mds(desc);
2884 		mirror->pg_recoalesce = 1;
2885 	}
2886 	hdr->completion_ops->completion(hdr);
2887 }
2888 
2889 static enum pnfs_try_status
2890 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2891 			const struct rpc_call_ops *call_ops,
2892 			struct pnfs_layout_segment *lseg,
2893 			int how)
2894 {
2895 	struct inode *inode = hdr->inode;
2896 	enum pnfs_try_status trypnfs;
2897 	struct nfs_server *nfss = NFS_SERVER(inode);
2898 
2899 	hdr->mds_ops = call_ops;
2900 
2901 	dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2902 		inode->i_ino, hdr->args.count, hdr->args.offset, how);
2903 	trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2904 	if (trypnfs != PNFS_NOT_ATTEMPTED)
2905 		nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2906 	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2907 	return trypnfs;
2908 }
2909 
2910 static void
2911 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2912 	      struct nfs_pgio_header *hdr, int how)
2913 {
2914 	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2915 	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2916 	enum pnfs_try_status trypnfs;
2917 
2918 	trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2919 	switch (trypnfs) {
2920 	case PNFS_NOT_ATTEMPTED:
2921 		pnfs_write_through_mds(desc, hdr);
2922 		break;
2923 	case PNFS_ATTEMPTED:
2924 		break;
2925 	case PNFS_TRY_AGAIN:
2926 		/* cleanup hdr and prepare to redo pnfs */
2927 		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2928 			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2929 			list_splice_init(&hdr->pages, &mirror->pg_list);
2930 			mirror->pg_recoalesce = 1;
2931 		}
2932 		hdr->mds_ops->rpc_release(hdr);
2933 	}
2934 }
2935 
2936 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2937 {
2938 	pnfs_put_lseg(hdr->lseg);
2939 	nfs_pgio_header_free(hdr);
2940 }
2941 
2942 int
2943 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2944 {
2945 	struct nfs_pgio_header *hdr;
2946 	int ret;
2947 
2948 	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2949 	if (!hdr) {
2950 		desc->pg_error = -ENOMEM;
2951 		return desc->pg_error;
2952 	}
2953 	nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2954 
2955 	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2956 	ret = nfs_generic_pgio(desc, hdr);
2957 	if (!ret)
2958 		pnfs_do_write(desc, hdr, desc->pg_ioflags);
2959 
2960 	return ret;
2961 }
2962 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2963 
2964 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2965 {
2966 	struct nfs_pageio_descriptor pgio;
2967 
2968 	/* Resend all requests through the MDS */
2969 	nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2970 	return nfs_pageio_resend(&pgio, hdr);
2971 }
2972 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2973 
2974 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2975 {
2976 	dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2977 	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2978 	    PNFS_LAYOUTRET_ON_ERROR) {
2979 		pnfs_return_layout(hdr->inode);
2980 	}
2981 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2982 		hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2983 }
2984 
2985 /*
2986  * Called by non rpc-based layout drivers
2987  */
2988 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2989 {
2990 	if (likely(!hdr->pnfs_error))
2991 		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2992 	trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2993 	if (unlikely(hdr->pnfs_error))
2994 		pnfs_ld_handle_read_error(hdr);
2995 	hdr->mds_ops->rpc_release(hdr);
2996 }
2997 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2998 
2999 static void
3000 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
3001 		struct nfs_pgio_header *hdr)
3002 {
3003 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3004 
3005 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3006 		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
3007 		nfs_pageio_reset_read_mds(desc);
3008 		mirror->pg_recoalesce = 1;
3009 	}
3010 	hdr->completion_ops->completion(hdr);
3011 }
3012 
3013 /*
3014  * Call the appropriate parallel I/O subsystem read function.
3015  */
3016 static enum pnfs_try_status
3017 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
3018 		       const struct rpc_call_ops *call_ops,
3019 		       struct pnfs_layout_segment *lseg)
3020 {
3021 	struct inode *inode = hdr->inode;
3022 	struct nfs_server *nfss = NFS_SERVER(inode);
3023 	enum pnfs_try_status trypnfs;
3024 
3025 	hdr->mds_ops = call_ops;
3026 
3027 	dprintk("%s: Reading ino:%lu %u@%llu\n",
3028 		__func__, inode->i_ino, hdr->args.count, hdr->args.offset);
3029 
3030 	trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
3031 	if (trypnfs != PNFS_NOT_ATTEMPTED)
3032 		nfs_inc_stats(inode, NFSIOS_PNFS_READ);
3033 	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3034 	return trypnfs;
3035 }
3036 
3037 /* Resend all requests through pnfs. */
3038 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
3039 			   unsigned int mirror_idx)
3040 {
3041 	struct nfs_pageio_descriptor pgio;
3042 
3043 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3044 		/* Prevent deadlocks with layoutreturn! */
3045 		pnfs_put_lseg(hdr->lseg);
3046 		hdr->lseg = NULL;
3047 
3048 		nfs_pageio_init_read(&pgio, hdr->inode, false,
3049 					hdr->completion_ops);
3050 		pgio.pg_mirror_idx = mirror_idx;
3051 		hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
3052 	}
3053 }
3054 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
3055 
3056 static void
3057 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
3058 {
3059 	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3060 	struct pnfs_layout_segment *lseg = desc->pg_lseg;
3061 	enum pnfs_try_status trypnfs;
3062 
3063 	trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3064 	switch (trypnfs) {
3065 	case PNFS_NOT_ATTEMPTED:
3066 		pnfs_read_through_mds(desc, hdr);
3067 		break;
3068 	case PNFS_ATTEMPTED:
3069 		break;
3070 	case PNFS_TRY_AGAIN:
3071 		/* cleanup hdr and prepare to redo pnfs */
3072 		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3073 			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3074 			list_splice_init(&hdr->pages, &mirror->pg_list);
3075 			mirror->pg_recoalesce = 1;
3076 		}
3077 		hdr->mds_ops->rpc_release(hdr);
3078 	}
3079 }
3080 
3081 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3082 {
3083 	pnfs_put_lseg(hdr->lseg);
3084 	nfs_pgio_header_free(hdr);
3085 }
3086 
3087 int
3088 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3089 {
3090 	struct nfs_pgio_header *hdr;
3091 	int ret;
3092 
3093 	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3094 	if (!hdr) {
3095 		desc->pg_error = -ENOMEM;
3096 		return desc->pg_error;
3097 	}
3098 	nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3099 	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3100 	ret = nfs_generic_pgio(desc, hdr);
3101 	if (!ret)
3102 		pnfs_do_read(desc, hdr);
3103 	return ret;
3104 }
3105 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3106 
3107 static void pnfs_clear_layoutcommitting(struct inode *inode)
3108 {
3109 	unsigned long *bitlock = &NFS_I(inode)->flags;
3110 
3111 	clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3112 	smp_mb__after_atomic();
3113 	wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3114 }
3115 
3116 /*
3117  * There can be multiple RW segments.
3118  */
3119 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3120 {
3121 	struct pnfs_layout_segment *lseg;
3122 
3123 	list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3124 		if (lseg->pls_range.iomode == IOMODE_RW &&
3125 		    test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3126 			list_add(&lseg->pls_lc_list, listp);
3127 	}
3128 }
3129 
3130 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3131 {
3132 	struct pnfs_layout_segment *lseg, *tmp;
3133 
3134 	/* Matched by references in pnfs_set_layoutcommit */
3135 	list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3136 		list_del_init(&lseg->pls_lc_list);
3137 		pnfs_put_lseg(lseg);
3138 	}
3139 
3140 	pnfs_clear_layoutcommitting(inode);
3141 }
3142 
3143 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3144 {
3145 	pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3146 }
3147 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3148 
3149 void
3150 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3151 		loff_t end_pos)
3152 {
3153 	struct nfs_inode *nfsi = NFS_I(inode);
3154 	bool mark_as_dirty = false;
3155 
3156 	spin_lock(&inode->i_lock);
3157 	if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3158 		nfsi->layout->plh_lwb = end_pos;
3159 		mark_as_dirty = true;
3160 		dprintk("%s: Set layoutcommit for inode %lu ",
3161 			__func__, inode->i_ino);
3162 	} else if (end_pos > nfsi->layout->plh_lwb)
3163 		nfsi->layout->plh_lwb = end_pos;
3164 	if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3165 		/* references matched in nfs4_layoutcommit_release */
3166 		pnfs_get_lseg(lseg);
3167 	}
3168 	spin_unlock(&inode->i_lock);
3169 	dprintk("%s: lseg %p end_pos %llu\n",
3170 		__func__, lseg, nfsi->layout->plh_lwb);
3171 
3172 	/* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3173 	 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3174 	if (mark_as_dirty)
3175 		mark_inode_dirty_sync(inode);
3176 }
3177 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3178 
3179 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3180 {
3181 	struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3182 
3183 	if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3184 		nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3185 	pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3186 }
3187 
3188 /*
3189  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3190  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3191  * data to disk to allow the server to recover the data if it crashes.
3192  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3193  * is off, and a COMMIT is sent to a data server, or
3194  * if WRITEs to a data server return NFS_DATA_SYNC.
3195  */
3196 int
3197 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3198 {
3199 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3200 	struct nfs4_layoutcommit_data *data;
3201 	struct nfs_inode *nfsi = NFS_I(inode);
3202 	loff_t end_pos;
3203 	int status;
3204 
3205 	if (!pnfs_layoutcommit_outstanding(inode))
3206 		return 0;
3207 
3208 	dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3209 
3210 	status = -EAGAIN;
3211 	if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3212 		if (!sync)
3213 			goto out;
3214 		status = wait_on_bit_lock_action(&nfsi->flags,
3215 				NFS_INO_LAYOUTCOMMITTING,
3216 				nfs_wait_bit_killable,
3217 				TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
3218 		if (status)
3219 			goto out;
3220 	}
3221 
3222 	status = -ENOMEM;
3223 	/* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3224 	data = kzalloc(sizeof(*data), nfs_io_gfp_mask());
3225 	if (!data)
3226 		goto clear_layoutcommitting;
3227 
3228 	status = 0;
3229 	spin_lock(&inode->i_lock);
3230 	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3231 		goto out_unlock;
3232 
3233 	INIT_LIST_HEAD(&data->lseg_list);
3234 	pnfs_list_write_lseg(inode, &data->lseg_list);
3235 
3236 	end_pos = nfsi->layout->plh_lwb;
3237 
3238 	nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3239 	data->cred = get_cred(nfsi->layout->plh_lc_cred);
3240 	spin_unlock(&inode->i_lock);
3241 
3242 	data->args.inode = inode;
3243 	nfs_fattr_init(&data->fattr);
3244 	data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3245 	data->res.fattr = &data->fattr;
3246 	if (end_pos != 0)
3247 		data->args.lastbytewritten = end_pos - 1;
3248 	else
3249 		data->args.lastbytewritten = U64_MAX;
3250 	data->res.server = NFS_SERVER(inode);
3251 
3252 	if (ld->prepare_layoutcommit) {
3253 		status = ld->prepare_layoutcommit(&data->args);
3254 		if (status) {
3255 			put_cred(data->cred);
3256 			spin_lock(&inode->i_lock);
3257 			set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3258 			if (end_pos > nfsi->layout->plh_lwb)
3259 				nfsi->layout->plh_lwb = end_pos;
3260 			goto out_unlock;
3261 		}
3262 	}
3263 
3264 
3265 	status = nfs4_proc_layoutcommit(data, sync);
3266 out:
3267 	if (status)
3268 		mark_inode_dirty_sync(inode);
3269 	dprintk("<-- %s status %d\n", __func__, status);
3270 	return status;
3271 out_unlock:
3272 	spin_unlock(&inode->i_lock);
3273 	kfree(data);
3274 clear_layoutcommitting:
3275 	pnfs_clear_layoutcommitting(inode);
3276 	goto out;
3277 }
3278 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3279 
3280 int
3281 pnfs_generic_sync(struct inode *inode, bool datasync)
3282 {
3283 	return pnfs_layoutcommit_inode(inode, true);
3284 }
3285 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3286 
3287 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3288 {
3289 	struct nfs4_threshold *thp;
3290 
3291 	thp = kzalloc(sizeof(*thp), nfs_io_gfp_mask());
3292 	if (!thp) {
3293 		dprintk("%s mdsthreshold allocation failed\n", __func__);
3294 		return NULL;
3295 	}
3296 	return thp;
3297 }
3298 
3299 #if IS_ENABLED(CONFIG_NFS_V4_2)
3300 int
3301 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3302 {
3303 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3304 	struct nfs_server *server = NFS_SERVER(inode);
3305 	struct nfs_inode *nfsi = NFS_I(inode);
3306 	struct nfs42_layoutstat_data *data;
3307 	struct pnfs_layout_hdr *hdr;
3308 	int status = 0;
3309 
3310 	if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3311 		goto out;
3312 
3313 	if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3314 		goto out;
3315 
3316 	if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3317 		goto out;
3318 
3319 	spin_lock(&inode->i_lock);
3320 	if (!NFS_I(inode)->layout) {
3321 		spin_unlock(&inode->i_lock);
3322 		goto out_clear_layoutstats;
3323 	}
3324 	hdr = NFS_I(inode)->layout;
3325 	pnfs_get_layout_hdr(hdr);
3326 	spin_unlock(&inode->i_lock);
3327 
3328 	data = kzalloc(sizeof(*data), gfp_flags);
3329 	if (!data) {
3330 		status = -ENOMEM;
3331 		goto out_put;
3332 	}
3333 
3334 	data->args.fh = NFS_FH(inode);
3335 	data->args.inode = inode;
3336 	status = ld->prepare_layoutstats(&data->args);
3337 	if (status)
3338 		goto out_free;
3339 
3340 	status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3341 
3342 out:
3343 	dprintk("%s returns %d\n", __func__, status);
3344 	return status;
3345 
3346 out_free:
3347 	kfree(data);
3348 out_put:
3349 	pnfs_put_layout_hdr(hdr);
3350 out_clear_layoutstats:
3351 	smp_mb__before_atomic();
3352 	clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3353 	smp_mb__after_atomic();
3354 	goto out;
3355 }
3356 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3357 #endif
3358 
3359 unsigned int layoutstats_timer;
3360 module_param(layoutstats_timer, uint, 0644);
3361 EXPORT_SYMBOL_GPL(layoutstats_timer);
3362