xref: /openbmc/linux/fs/nfs/pnfs.c (revision f7d84fa7)
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 
41 #define NFSDBG_FACILITY		NFSDBG_PNFS
42 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
43 
44 /* Locking:
45  *
46  * pnfs_spinlock:
47  *      protects pnfs_modules_tbl.
48  */
49 static DEFINE_SPINLOCK(pnfs_spinlock);
50 
51 /*
52  * pnfs_modules_tbl holds all pnfs modules
53  */
54 static LIST_HEAD(pnfs_modules_tbl);
55 
56 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
57 static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
58 		struct list_head *free_me,
59 		const struct pnfs_layout_range *range,
60 		u32 seq);
61 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
62 		                struct list_head *tmp_list);
63 
64 /* Return the registered pnfs layout driver module matching given id */
65 static struct pnfs_layoutdriver_type *
66 find_pnfs_driver_locked(u32 id)
67 {
68 	struct pnfs_layoutdriver_type *local;
69 
70 	list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
71 		if (local->id == id)
72 			goto out;
73 	local = NULL;
74 out:
75 	dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
76 	return local;
77 }
78 
79 static struct pnfs_layoutdriver_type *
80 find_pnfs_driver(u32 id)
81 {
82 	struct pnfs_layoutdriver_type *local;
83 
84 	spin_lock(&pnfs_spinlock);
85 	local = find_pnfs_driver_locked(id);
86 	if (local != NULL && !try_module_get(local->owner)) {
87 		dprintk("%s: Could not grab reference on module\n", __func__);
88 		local = NULL;
89 	}
90 	spin_unlock(&pnfs_spinlock);
91 	return local;
92 }
93 
94 void
95 unset_pnfs_layoutdriver(struct nfs_server *nfss)
96 {
97 	if (nfss->pnfs_curr_ld) {
98 		if (nfss->pnfs_curr_ld->clear_layoutdriver)
99 			nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
100 		/* Decrement the MDS count. Purge the deviceid cache if zero */
101 		if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
102 			nfs4_deviceid_purge_client(nfss->nfs_client);
103 		module_put(nfss->pnfs_curr_ld->owner);
104 	}
105 	nfss->pnfs_curr_ld = NULL;
106 }
107 
108 /*
109  * When the server sends a list of layout types, we choose one in the order
110  * given in the list below.
111  *
112  * FIXME: should this list be configurable in some fashion? module param?
113  * 	  mount option? something else?
114  */
115 static const u32 ld_prefs[] = {
116 	LAYOUT_SCSI,
117 	LAYOUT_BLOCK_VOLUME,
118 	LAYOUT_OSD2_OBJECTS,
119 	LAYOUT_FLEX_FILES,
120 	LAYOUT_NFSV4_1_FILES,
121 	0
122 };
123 
124 static int
125 ld_cmp(const void *e1, const void *e2)
126 {
127 	u32 ld1 = *((u32 *)e1);
128 	u32 ld2 = *((u32 *)e2);
129 	int i;
130 
131 	for (i = 0; ld_prefs[i] != 0; i++) {
132 		if (ld1 == ld_prefs[i])
133 			return -1;
134 
135 		if (ld2 == ld_prefs[i])
136 			return 1;
137 	}
138 	return 0;
139 }
140 
141 /*
142  * Try to set the server's pnfs module to the pnfs layout type specified by id.
143  * Currently only one pNFS layout driver per filesystem is supported.
144  *
145  * @ids array of layout types supported by MDS.
146  */
147 void
148 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
149 		      struct nfs_fsinfo *fsinfo)
150 {
151 	struct pnfs_layoutdriver_type *ld_type = NULL;
152 	u32 id;
153 	int i;
154 
155 	if (fsinfo->nlayouttypes == 0)
156 		goto out_no_driver;
157 	if (!(server->nfs_client->cl_exchange_flags &
158 		 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
159 		printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
160 			__func__, server->nfs_client->cl_exchange_flags);
161 		goto out_no_driver;
162 	}
163 
164 	sort(fsinfo->layouttype, fsinfo->nlayouttypes,
165 		sizeof(*fsinfo->layouttype), ld_cmp, NULL);
166 
167 	for (i = 0; i < fsinfo->nlayouttypes; i++) {
168 		id = fsinfo->layouttype[i];
169 		ld_type = find_pnfs_driver(id);
170 		if (!ld_type) {
171 			request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
172 					id);
173 			ld_type = find_pnfs_driver(id);
174 		}
175 		if (ld_type)
176 			break;
177 	}
178 
179 	if (!ld_type) {
180 		dprintk("%s: No pNFS module found!\n", __func__);
181 		goto out_no_driver;
182 	}
183 
184 	server->pnfs_curr_ld = ld_type;
185 	if (ld_type->set_layoutdriver
186 	    && ld_type->set_layoutdriver(server, mntfh)) {
187 		printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
188 			"driver %u.\n", __func__, id);
189 		module_put(ld_type->owner);
190 		goto out_no_driver;
191 	}
192 	/* Bump the MDS count */
193 	atomic_inc(&server->nfs_client->cl_mds_count);
194 
195 	dprintk("%s: pNFS module for %u set\n", __func__, id);
196 	return;
197 
198 out_no_driver:
199 	dprintk("%s: Using NFSv4 I/O\n", __func__);
200 	server->pnfs_curr_ld = NULL;
201 }
202 
203 int
204 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
205 {
206 	int status = -EINVAL;
207 	struct pnfs_layoutdriver_type *tmp;
208 
209 	if (ld_type->id == 0) {
210 		printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
211 		return status;
212 	}
213 	if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
214 		printk(KERN_ERR "NFS: %s Layout driver must provide "
215 		       "alloc_lseg and free_lseg.\n", __func__);
216 		return status;
217 	}
218 
219 	spin_lock(&pnfs_spinlock);
220 	tmp = find_pnfs_driver_locked(ld_type->id);
221 	if (!tmp) {
222 		list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
223 		status = 0;
224 		dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
225 			ld_type->name);
226 	} else {
227 		printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
228 			__func__, ld_type->id);
229 	}
230 	spin_unlock(&pnfs_spinlock);
231 
232 	return status;
233 }
234 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
235 
236 void
237 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
238 {
239 	dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
240 	spin_lock(&pnfs_spinlock);
241 	list_del(&ld_type->pnfs_tblid);
242 	spin_unlock(&pnfs_spinlock);
243 }
244 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
245 
246 /*
247  * pNFS client layout cache
248  */
249 
250 /* Need to hold i_lock if caller does not already hold reference */
251 void
252 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
253 {
254 	atomic_inc(&lo->plh_refcount);
255 }
256 
257 static struct pnfs_layout_hdr *
258 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
259 {
260 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
261 	return ld->alloc_layout_hdr(ino, gfp_flags);
262 }
263 
264 static void
265 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
266 {
267 	struct nfs_server *server = NFS_SERVER(lo->plh_inode);
268 	struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
269 
270 	if (!list_empty(&lo->plh_layouts)) {
271 		struct nfs_client *clp = server->nfs_client;
272 
273 		spin_lock(&clp->cl_lock);
274 		list_del_init(&lo->plh_layouts);
275 		spin_unlock(&clp->cl_lock);
276 	}
277 	put_rpccred(lo->plh_lc_cred);
278 	return ld->free_layout_hdr(lo);
279 }
280 
281 static void
282 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
283 {
284 	struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
285 	dprintk("%s: freeing layout cache %p\n", __func__, lo);
286 	nfsi->layout = NULL;
287 	/* Reset MDS Threshold I/O counters */
288 	nfsi->write_io = 0;
289 	nfsi->read_io = 0;
290 }
291 
292 void
293 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
294 {
295 	struct inode *inode = lo->plh_inode;
296 
297 	pnfs_layoutreturn_before_put_layout_hdr(lo);
298 
299 	if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
300 		if (!list_empty(&lo->plh_segs))
301 			WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
302 		pnfs_detach_layout_hdr(lo);
303 		spin_unlock(&inode->i_lock);
304 		pnfs_free_layout_hdr(lo);
305 	}
306 }
307 
308 static void
309 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
310 			 u32 seq)
311 {
312 	if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
313 		iomode = IOMODE_ANY;
314 	lo->plh_return_iomode = iomode;
315 	set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
316 	if (seq != 0) {
317 		WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
318 		lo->plh_return_seq = seq;
319 	}
320 }
321 
322 static void
323 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
324 {
325 	struct pnfs_layout_segment *lseg;
326 	lo->plh_return_iomode = 0;
327 	lo->plh_return_seq = 0;
328 	clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
329 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
330 		if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
331 			continue;
332 		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
333 	}
334 }
335 
336 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
337 {
338 	clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
339 	clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
340 	smp_mb__after_atomic();
341 	wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
342 	rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
343 }
344 
345 static void
346 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
347 		struct list_head *free_me)
348 {
349 	clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
350 	clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
351 	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
352 		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
353 	if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
354 		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
355 }
356 
357 /*
358  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
359  *
360  * In order to continue using the pnfs_layout_hdr, a full recovery
361  * is required.
362  * Note that caller must hold inode->i_lock.
363  */
364 int
365 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
366 		struct list_head *lseg_list)
367 {
368 	struct pnfs_layout_range range = {
369 		.iomode = IOMODE_ANY,
370 		.offset = 0,
371 		.length = NFS4_MAX_UINT64,
372 	};
373 	struct pnfs_layout_segment *lseg, *next;
374 
375 	set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
376 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
377 		pnfs_clear_lseg_state(lseg, lseg_list);
378 	pnfs_clear_layoutreturn_info(lo);
379 	pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
380 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
381 	    !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
382 		pnfs_clear_layoutreturn_waitbit(lo);
383 	return !list_empty(&lo->plh_segs);
384 }
385 
386 static int
387 pnfs_iomode_to_fail_bit(u32 iomode)
388 {
389 	return iomode == IOMODE_RW ?
390 		NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
391 }
392 
393 static void
394 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
395 {
396 	lo->plh_retry_timestamp = jiffies;
397 	if (!test_and_set_bit(fail_bit, &lo->plh_flags))
398 		atomic_inc(&lo->plh_refcount);
399 }
400 
401 static void
402 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
403 {
404 	if (test_and_clear_bit(fail_bit, &lo->plh_flags))
405 		atomic_dec(&lo->plh_refcount);
406 }
407 
408 static void
409 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
410 {
411 	struct inode *inode = lo->plh_inode;
412 	struct pnfs_layout_range range = {
413 		.iomode = iomode,
414 		.offset = 0,
415 		.length = NFS4_MAX_UINT64,
416 	};
417 	LIST_HEAD(head);
418 
419 	spin_lock(&inode->i_lock);
420 	pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
421 	pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
422 	spin_unlock(&inode->i_lock);
423 	pnfs_free_lseg_list(&head);
424 	dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
425 			iomode == IOMODE_RW ?  "RW" : "READ");
426 }
427 
428 static bool
429 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
430 {
431 	unsigned long start, end;
432 	int fail_bit = pnfs_iomode_to_fail_bit(iomode);
433 
434 	if (test_bit(fail_bit, &lo->plh_flags) == 0)
435 		return false;
436 	end = jiffies;
437 	start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
438 	if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
439 		/* It is time to retry the failed layoutgets */
440 		pnfs_layout_clear_fail_bit(lo, fail_bit);
441 		return false;
442 	}
443 	return true;
444 }
445 
446 static void
447 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
448 		const struct pnfs_layout_range *range,
449 		const nfs4_stateid *stateid)
450 {
451 	INIT_LIST_HEAD(&lseg->pls_list);
452 	INIT_LIST_HEAD(&lseg->pls_lc_list);
453 	atomic_set(&lseg->pls_refcount, 1);
454 	set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
455 	lseg->pls_layout = lo;
456 	lseg->pls_range = *range;
457 	lseg->pls_seq = be32_to_cpu(stateid->seqid);
458 }
459 
460 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
461 {
462 	if (lseg != NULL) {
463 		struct inode *inode = lseg->pls_layout->plh_inode;
464 		NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
465 	}
466 }
467 
468 static void
469 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
470 		struct pnfs_layout_segment *lseg)
471 {
472 	WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
473 	list_del_init(&lseg->pls_list);
474 	/* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
475 	atomic_dec(&lo->plh_refcount);
476 	if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
477 		return;
478 	if (list_empty(&lo->plh_segs) &&
479 	    !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
480 	    !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
481 		if (atomic_read(&lo->plh_outstanding) == 0)
482 			set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
483 		clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
484 	}
485 }
486 
487 static bool
488 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
489 		struct pnfs_layout_segment *lseg)
490 {
491 	if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
492 	    pnfs_layout_is_valid(lo)) {
493 		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
494 		list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
495 		return true;
496 	}
497 	return false;
498 }
499 
500 void
501 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
502 {
503 	struct pnfs_layout_hdr *lo;
504 	struct inode *inode;
505 
506 	if (!lseg)
507 		return;
508 
509 	dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
510 		atomic_read(&lseg->pls_refcount),
511 		test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
512 
513 	lo = lseg->pls_layout;
514 	inode = lo->plh_inode;
515 
516 	if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
517 		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
518 			spin_unlock(&inode->i_lock);
519 			return;
520 		}
521 		pnfs_get_layout_hdr(lo);
522 		pnfs_layout_remove_lseg(lo, lseg);
523 		if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
524 			lseg = NULL;
525 		spin_unlock(&inode->i_lock);
526 		pnfs_free_lseg(lseg);
527 		pnfs_put_layout_hdr(lo);
528 	}
529 }
530 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
531 
532 static void pnfs_free_lseg_async_work(struct work_struct *work)
533 {
534 	struct pnfs_layout_segment *lseg;
535 	struct pnfs_layout_hdr *lo;
536 
537 	lseg = container_of(work, struct pnfs_layout_segment, pls_work);
538 	lo = lseg->pls_layout;
539 
540 	pnfs_free_lseg(lseg);
541 	pnfs_put_layout_hdr(lo);
542 }
543 
544 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
545 {
546 	INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
547 	schedule_work(&lseg->pls_work);
548 }
549 
550 void
551 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
552 {
553 	if (!lseg)
554 		return;
555 
556 	assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
557 
558 	dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
559 		atomic_read(&lseg->pls_refcount),
560 		test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
561 	if (atomic_dec_and_test(&lseg->pls_refcount)) {
562 		struct pnfs_layout_hdr *lo = lseg->pls_layout;
563 		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
564 			return;
565 		pnfs_layout_remove_lseg(lo, lseg);
566 		if (!pnfs_cache_lseg_for_layoutreturn(lo, lseg)) {
567 			pnfs_get_layout_hdr(lo);
568 			pnfs_free_lseg_async(lseg);
569 		}
570 	}
571 }
572 
573 /*
574  * is l2 fully contained in l1?
575  *   start1                             end1
576  *   [----------------------------------)
577  *           start2           end2
578  *           [----------------)
579  */
580 static bool
581 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
582 		 const struct pnfs_layout_range *l2)
583 {
584 	u64 start1 = l1->offset;
585 	u64 end1 = pnfs_end_offset(start1, l1->length);
586 	u64 start2 = l2->offset;
587 	u64 end2 = pnfs_end_offset(start2, l2->length);
588 
589 	return (start1 <= start2) && (end1 >= end2);
590 }
591 
592 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
593 		struct list_head *tmp_list)
594 {
595 	if (!atomic_dec_and_test(&lseg->pls_refcount))
596 		return false;
597 	pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
598 	list_add(&lseg->pls_list, tmp_list);
599 	return true;
600 }
601 
602 /* Returns 1 if lseg is removed from list, 0 otherwise */
603 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
604 			     struct list_head *tmp_list)
605 {
606 	int rv = 0;
607 
608 	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
609 		/* Remove the reference keeping the lseg in the
610 		 * list.  It will now be removed when all
611 		 * outstanding io is finished.
612 		 */
613 		dprintk("%s: lseg %p ref %d\n", __func__, lseg,
614 			atomic_read(&lseg->pls_refcount));
615 		if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
616 			rv = 1;
617 	}
618 	return rv;
619 }
620 
621 /*
622  * Compare 2 layout stateid sequence ids, to see which is newer,
623  * taking into account wraparound issues.
624  */
625 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
626 {
627 	return (s32)(s1 - s2) > 0;
628 }
629 
630 static bool
631 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
632 		 const struct pnfs_layout_range *recall_range)
633 {
634 	return (recall_range->iomode == IOMODE_ANY ||
635 		lseg_range->iomode == recall_range->iomode) &&
636 	       pnfs_lseg_range_intersecting(lseg_range, recall_range);
637 }
638 
639 static bool
640 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
641 		const struct pnfs_layout_range *recall_range,
642 		u32 seq)
643 {
644 	if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
645 		return false;
646 	if (recall_range == NULL)
647 		return true;
648 	return pnfs_should_free_range(&lseg->pls_range, recall_range);
649 }
650 
651 /**
652  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
653  * @lo: layout header containing the lsegs
654  * @tmp_list: list head where doomed lsegs should go
655  * @recall_range: optional recall range argument to match (may be NULL)
656  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
657  *
658  * Walk the list of lsegs in the layout header, and tear down any that should
659  * be destroyed. If "recall_range" is specified then the segment must match
660  * that range. If "seq" is non-zero, then only match segments that were handed
661  * out at or before that sequence.
662  *
663  * Returns number of matching invalid lsegs remaining in list after scanning
664  * it and purging them.
665  */
666 int
667 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
668 			    struct list_head *tmp_list,
669 			    const struct pnfs_layout_range *recall_range,
670 			    u32 seq)
671 {
672 	struct pnfs_layout_segment *lseg, *next;
673 	int remaining = 0;
674 
675 	dprintk("%s:Begin lo %p\n", __func__, lo);
676 
677 	if (list_empty(&lo->plh_segs))
678 		return 0;
679 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
680 		if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
681 			dprintk("%s: freeing lseg %p iomode %d seq %u"
682 				"offset %llu length %llu\n", __func__,
683 				lseg, lseg->pls_range.iomode, lseg->pls_seq,
684 				lseg->pls_range.offset, lseg->pls_range.length);
685 			if (!mark_lseg_invalid(lseg, tmp_list))
686 				remaining++;
687 		}
688 	dprintk("%s:Return %i\n", __func__, remaining);
689 	return remaining;
690 }
691 
692 static void
693 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
694 		struct list_head *free_me,
695 		const struct pnfs_layout_range *range,
696 		u32 seq)
697 {
698 	struct pnfs_layout_segment *lseg, *next;
699 
700 	list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
701 		if (pnfs_match_lseg_recall(lseg, range, seq))
702 			list_move_tail(&lseg->pls_list, free_me);
703 	}
704 }
705 
706 /* note free_me must contain lsegs from a single layout_hdr */
707 void
708 pnfs_free_lseg_list(struct list_head *free_me)
709 {
710 	struct pnfs_layout_segment *lseg, *tmp;
711 
712 	if (list_empty(free_me))
713 		return;
714 
715 	list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
716 		list_del(&lseg->pls_list);
717 		pnfs_free_lseg(lseg);
718 	}
719 }
720 
721 void
722 pnfs_destroy_layout(struct nfs_inode *nfsi)
723 {
724 	struct pnfs_layout_hdr *lo;
725 	LIST_HEAD(tmp_list);
726 
727 	spin_lock(&nfsi->vfs_inode.i_lock);
728 	lo = nfsi->layout;
729 	if (lo) {
730 		pnfs_get_layout_hdr(lo);
731 		pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
732 		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
733 		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
734 		spin_unlock(&nfsi->vfs_inode.i_lock);
735 		pnfs_free_lseg_list(&tmp_list);
736 		nfs_commit_inode(&nfsi->vfs_inode, 0);
737 		pnfs_put_layout_hdr(lo);
738 	} else
739 		spin_unlock(&nfsi->vfs_inode.i_lock);
740 }
741 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
742 
743 static bool
744 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
745 		struct list_head *layout_list)
746 {
747 	struct pnfs_layout_hdr *lo;
748 	bool ret = false;
749 
750 	spin_lock(&inode->i_lock);
751 	lo = NFS_I(inode)->layout;
752 	if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
753 		pnfs_get_layout_hdr(lo);
754 		list_add(&lo->plh_bulk_destroy, layout_list);
755 		ret = true;
756 	}
757 	spin_unlock(&inode->i_lock);
758 	return ret;
759 }
760 
761 /* Caller must hold rcu_read_lock and clp->cl_lock */
762 static int
763 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
764 		struct nfs_server *server,
765 		struct list_head *layout_list)
766 {
767 	struct pnfs_layout_hdr *lo, *next;
768 	struct inode *inode;
769 
770 	list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
771 		if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
772 			continue;
773 		inode = igrab(lo->plh_inode);
774 		if (inode == NULL)
775 			continue;
776 		list_del_init(&lo->plh_layouts);
777 		if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
778 			continue;
779 		rcu_read_unlock();
780 		spin_unlock(&clp->cl_lock);
781 		iput(inode);
782 		spin_lock(&clp->cl_lock);
783 		rcu_read_lock();
784 		return -EAGAIN;
785 	}
786 	return 0;
787 }
788 
789 static int
790 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
791 		bool is_bulk_recall)
792 {
793 	struct pnfs_layout_hdr *lo;
794 	struct inode *inode;
795 	LIST_HEAD(lseg_list);
796 	int ret = 0;
797 
798 	while (!list_empty(layout_list)) {
799 		lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
800 				plh_bulk_destroy);
801 		dprintk("%s freeing layout for inode %lu\n", __func__,
802 			lo->plh_inode->i_ino);
803 		inode = lo->plh_inode;
804 
805 		pnfs_layoutcommit_inode(inode, false);
806 
807 		spin_lock(&inode->i_lock);
808 		list_del_init(&lo->plh_bulk_destroy);
809 		if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
810 			if (is_bulk_recall)
811 				set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
812 			ret = -EAGAIN;
813 		}
814 		spin_unlock(&inode->i_lock);
815 		pnfs_free_lseg_list(&lseg_list);
816 		/* Free all lsegs that are attached to commit buckets */
817 		nfs_commit_inode(inode, 0);
818 		pnfs_put_layout_hdr(lo);
819 		iput(inode);
820 	}
821 	return ret;
822 }
823 
824 int
825 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
826 		struct nfs_fsid *fsid,
827 		bool is_recall)
828 {
829 	struct nfs_server *server;
830 	LIST_HEAD(layout_list);
831 
832 	spin_lock(&clp->cl_lock);
833 	rcu_read_lock();
834 restart:
835 	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
836 		if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
837 			continue;
838 		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
839 				server,
840 				&layout_list) != 0)
841 			goto restart;
842 	}
843 	rcu_read_unlock();
844 	spin_unlock(&clp->cl_lock);
845 
846 	if (list_empty(&layout_list))
847 		return 0;
848 	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
849 }
850 
851 int
852 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
853 		bool is_recall)
854 {
855 	struct nfs_server *server;
856 	LIST_HEAD(layout_list);
857 
858 	spin_lock(&clp->cl_lock);
859 	rcu_read_lock();
860 restart:
861 	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
862 		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
863 					server,
864 					&layout_list) != 0)
865 			goto restart;
866 	}
867 	rcu_read_unlock();
868 	spin_unlock(&clp->cl_lock);
869 
870 	if (list_empty(&layout_list))
871 		return 0;
872 	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
873 }
874 
875 /*
876  * Called by the state manger to remove all layouts established under an
877  * expired lease.
878  */
879 void
880 pnfs_destroy_all_layouts(struct nfs_client *clp)
881 {
882 	nfs4_deviceid_mark_client_invalid(clp);
883 	nfs4_deviceid_purge_client(clp);
884 
885 	pnfs_destroy_layouts_byclid(clp, false);
886 }
887 
888 /* update lo->plh_stateid with new if is more recent */
889 void
890 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
891 			bool update_barrier)
892 {
893 	u32 oldseq, newseq, new_barrier = 0;
894 
895 	oldseq = be32_to_cpu(lo->plh_stateid.seqid);
896 	newseq = be32_to_cpu(new->seqid);
897 
898 	if (!pnfs_layout_is_valid(lo)) {
899 		nfs4_stateid_copy(&lo->plh_stateid, new);
900 		lo->plh_barrier = newseq;
901 		pnfs_clear_layoutreturn_info(lo);
902 		clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
903 		return;
904 	}
905 	if (pnfs_seqid_is_newer(newseq, oldseq)) {
906 		nfs4_stateid_copy(&lo->plh_stateid, new);
907 		/*
908 		 * Because of wraparound, we want to keep the barrier
909 		 * "close" to the current seqids.
910 		 */
911 		new_barrier = newseq - atomic_read(&lo->plh_outstanding);
912 	}
913 	if (update_barrier)
914 		new_barrier = be32_to_cpu(new->seqid);
915 	else if (new_barrier == 0)
916 		return;
917 	if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
918 		lo->plh_barrier = new_barrier;
919 }
920 
921 static bool
922 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
923 		const nfs4_stateid *stateid)
924 {
925 	u32 seqid = be32_to_cpu(stateid->seqid);
926 
927 	return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
928 }
929 
930 /* lget is set to 1 if called from inside send_layoutget call chain */
931 static bool
932 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
933 {
934 	return lo->plh_block_lgets ||
935 		test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
936 }
937 
938 /*
939  * Get layout from server.
940  *    for now, assume that whole file layouts are requested.
941  *    arg->offset: 0
942  *    arg->length: all ones
943  */
944 static struct pnfs_layout_segment *
945 send_layoutget(struct pnfs_layout_hdr *lo,
946 	   struct nfs_open_context *ctx,
947 	   nfs4_stateid *stateid,
948 	   const struct pnfs_layout_range *range,
949 	   long *timeout, gfp_t gfp_flags)
950 {
951 	struct inode *ino = lo->plh_inode;
952 	struct nfs_server *server = NFS_SERVER(ino);
953 	struct nfs4_layoutget *lgp;
954 	loff_t i_size;
955 
956 	dprintk("--> %s\n", __func__);
957 
958 	/*
959 	 * Synchronously retrieve layout information from server and
960 	 * store in lseg. If we race with a concurrent seqid morphing
961 	 * op, then re-send the LAYOUTGET.
962 	 */
963 	lgp = kzalloc(sizeof(*lgp), gfp_flags);
964 	if (lgp == NULL)
965 		return ERR_PTR(-ENOMEM);
966 
967 	i_size = i_size_read(ino);
968 
969 	lgp->args.minlength = PAGE_SIZE;
970 	if (lgp->args.minlength > range->length)
971 		lgp->args.minlength = range->length;
972 	if (range->iomode == IOMODE_READ) {
973 		if (range->offset >= i_size)
974 			lgp->args.minlength = 0;
975 		else if (i_size - range->offset < lgp->args.minlength)
976 			lgp->args.minlength = i_size - range->offset;
977 	}
978 	lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
979 	pnfs_copy_range(&lgp->args.range, range);
980 	lgp->args.type = server->pnfs_curr_ld->id;
981 	lgp->args.inode = ino;
982 	lgp->args.ctx = get_nfs_open_context(ctx);
983 	nfs4_stateid_copy(&lgp->args.stateid, stateid);
984 	lgp->gfp_flags = gfp_flags;
985 	lgp->cred = lo->plh_lc_cred;
986 
987 	return nfs4_proc_layoutget(lgp, timeout, gfp_flags);
988 }
989 
990 static void pnfs_clear_layoutcommit(struct inode *inode,
991 		struct list_head *head)
992 {
993 	struct nfs_inode *nfsi = NFS_I(inode);
994 	struct pnfs_layout_segment *lseg, *tmp;
995 
996 	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
997 		return;
998 	list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
999 		if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1000 			continue;
1001 		pnfs_lseg_dec_and_remove_zero(lseg, head);
1002 	}
1003 }
1004 
1005 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1006 		const nfs4_stateid *arg_stateid,
1007 		const struct pnfs_layout_range *range,
1008 		const nfs4_stateid *stateid)
1009 {
1010 	struct inode *inode = lo->plh_inode;
1011 	LIST_HEAD(freeme);
1012 
1013 	spin_lock(&inode->i_lock);
1014 	if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
1015 	    !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1016 		goto out_unlock;
1017 	if (stateid) {
1018 		u32 seq = be32_to_cpu(arg_stateid->seqid);
1019 
1020 		pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1021 		pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1022 		pnfs_set_layout_stateid(lo, stateid, true);
1023 	} else
1024 		pnfs_mark_layout_stateid_invalid(lo, &freeme);
1025 out_unlock:
1026 	pnfs_clear_layoutreturn_waitbit(lo);
1027 	spin_unlock(&inode->i_lock);
1028 	pnfs_free_lseg_list(&freeme);
1029 
1030 }
1031 
1032 static bool
1033 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1034 		nfs4_stateid *stateid,
1035 		enum pnfs_iomode *iomode)
1036 {
1037 	/* Serialise LAYOUTGET/LAYOUTRETURN */
1038 	if (atomic_read(&lo->plh_outstanding) != 0)
1039 		return false;
1040 	if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1041 		return false;
1042 	set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1043 	pnfs_get_layout_hdr(lo);
1044 	if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1045 		if (stateid != NULL) {
1046 			nfs4_stateid_copy(stateid, &lo->plh_stateid);
1047 			if (lo->plh_return_seq != 0)
1048 				stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1049 		}
1050 		if (iomode != NULL)
1051 			*iomode = lo->plh_return_iomode;
1052 		pnfs_clear_layoutreturn_info(lo);
1053 		return true;
1054 	}
1055 	if (stateid != NULL)
1056 		nfs4_stateid_copy(stateid, &lo->plh_stateid);
1057 	if (iomode != NULL)
1058 		*iomode = IOMODE_ANY;
1059 	return true;
1060 }
1061 
1062 static void
1063 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1064 		struct pnfs_layout_hdr *lo,
1065 		const nfs4_stateid *stateid,
1066 		enum pnfs_iomode iomode)
1067 {
1068 	struct inode *inode = lo->plh_inode;
1069 
1070 	args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1071 	args->inode = inode;
1072 	args->range.iomode = iomode;
1073 	args->range.offset = 0;
1074 	args->range.length = NFS4_MAX_UINT64;
1075 	args->layout = lo;
1076 	nfs4_stateid_copy(&args->stateid, stateid);
1077 }
1078 
1079 static int
1080 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
1081 		       enum pnfs_iomode iomode, bool sync)
1082 {
1083 	struct inode *ino = lo->plh_inode;
1084 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1085 	struct nfs4_layoutreturn *lrp;
1086 	int status = 0;
1087 
1088 	lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1089 	if (unlikely(lrp == NULL)) {
1090 		status = -ENOMEM;
1091 		spin_lock(&ino->i_lock);
1092 		pnfs_clear_layoutreturn_waitbit(lo);
1093 		spin_unlock(&ino->i_lock);
1094 		pnfs_put_layout_hdr(lo);
1095 		goto out;
1096 	}
1097 
1098 	pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1099 	lrp->args.ld_private = &lrp->ld_private;
1100 	lrp->clp = NFS_SERVER(ino)->nfs_client;
1101 	lrp->cred = lo->plh_lc_cred;
1102 	if (ld->prepare_layoutreturn)
1103 		ld->prepare_layoutreturn(&lrp->args);
1104 
1105 	status = nfs4_proc_layoutreturn(lrp, sync);
1106 out:
1107 	dprintk("<-- %s status: %d\n", __func__, status);
1108 	return status;
1109 }
1110 
1111 /* Return true if layoutreturn is needed */
1112 static bool
1113 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1114 {
1115 	struct pnfs_layout_segment *s;
1116 
1117 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1118 		return false;
1119 
1120 	/* Defer layoutreturn until all lsegs are done */
1121 	list_for_each_entry(s, &lo->plh_segs, pls_list) {
1122 		if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1123 			return false;
1124 	}
1125 
1126 	return true;
1127 }
1128 
1129 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1130 {
1131 	struct inode *inode= lo->plh_inode;
1132 
1133 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1134 		return;
1135 	spin_lock(&inode->i_lock);
1136 	if (pnfs_layout_need_return(lo)) {
1137 		nfs4_stateid stateid;
1138 		enum pnfs_iomode iomode;
1139 		bool send;
1140 
1141 		send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1142 		spin_unlock(&inode->i_lock);
1143 		if (send) {
1144 			/* Send an async layoutreturn so we dont deadlock */
1145 			pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1146 		}
1147 	} else
1148 		spin_unlock(&inode->i_lock);
1149 }
1150 
1151 /*
1152  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1153  * when the layout segment list is empty.
1154  *
1155  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1156  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1157  * deviceid is marked invalid.
1158  */
1159 int
1160 _pnfs_return_layout(struct inode *ino)
1161 {
1162 	struct pnfs_layout_hdr *lo = NULL;
1163 	struct nfs_inode *nfsi = NFS_I(ino);
1164 	LIST_HEAD(tmp_list);
1165 	nfs4_stateid stateid;
1166 	int status = 0;
1167 	bool send;
1168 
1169 	dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1170 
1171 	spin_lock(&ino->i_lock);
1172 	lo = nfsi->layout;
1173 	if (!lo) {
1174 		spin_unlock(&ino->i_lock);
1175 		dprintk("NFS: %s no layout to return\n", __func__);
1176 		goto out;
1177 	}
1178 	/* Reference matched in nfs4_layoutreturn_release */
1179 	pnfs_get_layout_hdr(lo);
1180 	/* Is there an outstanding layoutreturn ? */
1181 	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1182 		spin_unlock(&ino->i_lock);
1183 		if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1184 					TASK_UNINTERRUPTIBLE))
1185 			goto out_put_layout_hdr;
1186 		spin_lock(&ino->i_lock);
1187 	}
1188 	pnfs_clear_layoutcommit(ino, &tmp_list);
1189 	pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1190 
1191 	if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1192 		struct pnfs_layout_range range = {
1193 			.iomode		= IOMODE_ANY,
1194 			.offset		= 0,
1195 			.length		= NFS4_MAX_UINT64,
1196 		};
1197 		NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1198 	}
1199 
1200 	/* Don't send a LAYOUTRETURN if list was initially empty */
1201 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1202 		spin_unlock(&ino->i_lock);
1203 		dprintk("NFS: %s no layout segments to return\n", __func__);
1204 		goto out_put_layout_hdr;
1205 	}
1206 
1207 	send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1208 	spin_unlock(&ino->i_lock);
1209 	if (send)
1210 		status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1211 out_put_layout_hdr:
1212 	pnfs_free_lseg_list(&tmp_list);
1213 	pnfs_put_layout_hdr(lo);
1214 out:
1215 	dprintk("<-- %s status: %d\n", __func__, status);
1216 	return status;
1217 }
1218 
1219 int
1220 pnfs_commit_and_return_layout(struct inode *inode)
1221 {
1222 	struct pnfs_layout_hdr *lo;
1223 	int ret;
1224 
1225 	spin_lock(&inode->i_lock);
1226 	lo = NFS_I(inode)->layout;
1227 	if (lo == NULL) {
1228 		spin_unlock(&inode->i_lock);
1229 		return 0;
1230 	}
1231 	pnfs_get_layout_hdr(lo);
1232 	/* Block new layoutgets and read/write to ds */
1233 	lo->plh_block_lgets++;
1234 	spin_unlock(&inode->i_lock);
1235 	filemap_fdatawait(inode->i_mapping);
1236 	ret = pnfs_layoutcommit_inode(inode, true);
1237 	if (ret == 0)
1238 		ret = _pnfs_return_layout(inode);
1239 	spin_lock(&inode->i_lock);
1240 	lo->plh_block_lgets--;
1241 	spin_unlock(&inode->i_lock);
1242 	pnfs_put_layout_hdr(lo);
1243 	return ret;
1244 }
1245 
1246 bool pnfs_roc(struct inode *ino,
1247 		struct nfs4_layoutreturn_args *args,
1248 		struct nfs4_layoutreturn_res *res,
1249 		const struct rpc_cred *cred)
1250 {
1251 	struct nfs_inode *nfsi = NFS_I(ino);
1252 	struct nfs_open_context *ctx;
1253 	struct nfs4_state *state;
1254 	struct pnfs_layout_hdr *lo;
1255 	struct pnfs_layout_segment *lseg, *next;
1256 	nfs4_stateid stateid;
1257 	enum pnfs_iomode iomode = 0;
1258 	bool layoutreturn = false, roc = false;
1259 	bool skip_read = false;
1260 
1261 	if (!nfs_have_layout(ino))
1262 		return false;
1263 retry:
1264 	spin_lock(&ino->i_lock);
1265 	lo = nfsi->layout;
1266 	if (!lo || !pnfs_layout_is_valid(lo) ||
1267 	    test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1268 		goto out_noroc;
1269 	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1270 		pnfs_get_layout_hdr(lo);
1271 		spin_unlock(&ino->i_lock);
1272 		wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1273 				TASK_UNINTERRUPTIBLE);
1274 		pnfs_put_layout_hdr(lo);
1275 		goto retry;
1276 	}
1277 
1278 	/* no roc if we hold a delegation */
1279 	if (nfs4_check_delegation(ino, FMODE_READ)) {
1280 		if (nfs4_check_delegation(ino, FMODE_WRITE))
1281 			goto out_noroc;
1282 		skip_read = true;
1283 	}
1284 
1285 	list_for_each_entry(ctx, &nfsi->open_files, list) {
1286 		state = ctx->state;
1287 		if (state == NULL)
1288 			continue;
1289 		/* Don't return layout if there is open file state */
1290 		if (state->state & FMODE_WRITE)
1291 			goto out_noroc;
1292 		if (state->state & FMODE_READ)
1293 			skip_read = true;
1294 	}
1295 
1296 
1297 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1298 		if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1299 			continue;
1300 		/* If we are sending layoutreturn, invalidate all valid lsegs */
1301 		if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1302 			continue;
1303 		/*
1304 		 * Note: mark lseg for return so pnfs_layout_remove_lseg
1305 		 * doesn't invalidate the layout for us.
1306 		 */
1307 		set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1308 		if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1309 			continue;
1310 		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1311 	}
1312 
1313 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1314 		goto out_noroc;
1315 
1316 	/* ROC in two conditions:
1317 	 * 1. there are ROC lsegs
1318 	 * 2. we don't send layoutreturn
1319 	 */
1320 	/* lo ref dropped in pnfs_roc_release() */
1321 	layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1322 	/* If the creds don't match, we can't compound the layoutreturn */
1323 	if (!layoutreturn || cred != lo->plh_lc_cred)
1324 		goto out_noroc;
1325 
1326 	roc = layoutreturn;
1327 	pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1328 	res->lrs_present = 0;
1329 	layoutreturn = false;
1330 
1331 out_noroc:
1332 	spin_unlock(&ino->i_lock);
1333 	pnfs_layoutcommit_inode(ino, true);
1334 	if (roc) {
1335 		struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1336 		if (ld->prepare_layoutreturn)
1337 			ld->prepare_layoutreturn(args);
1338 		return true;
1339 	}
1340 	if (layoutreturn)
1341 		pnfs_send_layoutreturn(lo, &stateid, iomode, true);
1342 	return false;
1343 }
1344 
1345 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1346 		struct nfs4_layoutreturn_res *res,
1347 		int ret)
1348 {
1349 	struct pnfs_layout_hdr *lo = args->layout;
1350 	const nfs4_stateid *arg_stateid = NULL;
1351 	const nfs4_stateid *res_stateid = NULL;
1352 	struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1353 
1354 	if (ret == 0) {
1355 		arg_stateid = &args->stateid;
1356 		if (res->lrs_present)
1357 			res_stateid = &res->stateid;
1358 	}
1359 	pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1360 			res_stateid);
1361 	if (ld_private && ld_private->ops && ld_private->ops->free)
1362 		ld_private->ops->free(ld_private);
1363 	pnfs_put_layout_hdr(lo);
1364 	trace_nfs4_layoutreturn_on_close(args->inode, 0);
1365 }
1366 
1367 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1368 {
1369 	struct nfs_inode *nfsi = NFS_I(ino);
1370         struct pnfs_layout_hdr *lo;
1371         bool sleep = false;
1372 
1373 	/* we might not have grabbed lo reference. so need to check under
1374 	 * i_lock */
1375         spin_lock(&ino->i_lock);
1376         lo = nfsi->layout;
1377         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1378                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1379                 sleep = true;
1380 	}
1381         spin_unlock(&ino->i_lock);
1382         return sleep;
1383 }
1384 
1385 /*
1386  * Compare two layout segments for sorting into layout cache.
1387  * We want to preferentially return RW over RO layouts, so ensure those
1388  * are seen first.
1389  */
1390 static s64
1391 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1392 	   const struct pnfs_layout_range *l2)
1393 {
1394 	s64 d;
1395 
1396 	/* high offset > low offset */
1397 	d = l1->offset - l2->offset;
1398 	if (d)
1399 		return d;
1400 
1401 	/* short length > long length */
1402 	d = l2->length - l1->length;
1403 	if (d)
1404 		return d;
1405 
1406 	/* read > read/write */
1407 	return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1408 }
1409 
1410 static bool
1411 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1412 		const struct pnfs_layout_range *l2)
1413 {
1414 	return pnfs_lseg_range_cmp(l1, l2) > 0;
1415 }
1416 
1417 static bool
1418 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1419 		struct pnfs_layout_segment *old)
1420 {
1421 	return false;
1422 }
1423 
1424 void
1425 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1426 		   struct pnfs_layout_segment *lseg,
1427 		   bool (*is_after)(const struct pnfs_layout_range *,
1428 			   const struct pnfs_layout_range *),
1429 		   bool (*do_merge)(struct pnfs_layout_segment *,
1430 			   struct pnfs_layout_segment *),
1431 		   struct list_head *free_me)
1432 {
1433 	struct pnfs_layout_segment *lp, *tmp;
1434 
1435 	dprintk("%s:Begin\n", __func__);
1436 
1437 	list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1438 		if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1439 			continue;
1440 		if (do_merge(lseg, lp)) {
1441 			mark_lseg_invalid(lp, free_me);
1442 			continue;
1443 		}
1444 		if (is_after(&lseg->pls_range, &lp->pls_range))
1445 			continue;
1446 		list_add_tail(&lseg->pls_list, &lp->pls_list);
1447 		dprintk("%s: inserted lseg %p "
1448 			"iomode %d offset %llu length %llu before "
1449 			"lp %p iomode %d offset %llu length %llu\n",
1450 			__func__, lseg, lseg->pls_range.iomode,
1451 			lseg->pls_range.offset, lseg->pls_range.length,
1452 			lp, lp->pls_range.iomode, lp->pls_range.offset,
1453 			lp->pls_range.length);
1454 		goto out;
1455 	}
1456 	list_add_tail(&lseg->pls_list, &lo->plh_segs);
1457 	dprintk("%s: inserted lseg %p "
1458 		"iomode %d offset %llu length %llu at tail\n",
1459 		__func__, lseg, lseg->pls_range.iomode,
1460 		lseg->pls_range.offset, lseg->pls_range.length);
1461 out:
1462 	pnfs_get_layout_hdr(lo);
1463 
1464 	dprintk("%s:Return\n", __func__);
1465 }
1466 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1467 
1468 static void
1469 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1470 		   struct pnfs_layout_segment *lseg,
1471 		   struct list_head *free_me)
1472 {
1473 	struct inode *inode = lo->plh_inode;
1474 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1475 
1476 	if (ld->add_lseg != NULL)
1477 		ld->add_lseg(lo, lseg, free_me);
1478 	else
1479 		pnfs_generic_layout_insert_lseg(lo, lseg,
1480 				pnfs_lseg_range_is_after,
1481 				pnfs_lseg_no_merge,
1482 				free_me);
1483 }
1484 
1485 static struct pnfs_layout_hdr *
1486 alloc_init_layout_hdr(struct inode *ino,
1487 		      struct nfs_open_context *ctx,
1488 		      gfp_t gfp_flags)
1489 {
1490 	struct pnfs_layout_hdr *lo;
1491 
1492 	lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1493 	if (!lo)
1494 		return NULL;
1495 	atomic_set(&lo->plh_refcount, 1);
1496 	INIT_LIST_HEAD(&lo->plh_layouts);
1497 	INIT_LIST_HEAD(&lo->plh_segs);
1498 	INIT_LIST_HEAD(&lo->plh_return_segs);
1499 	INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1500 	lo->plh_inode = ino;
1501 	lo->plh_lc_cred = get_rpccred(ctx->cred);
1502 	lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1503 	return lo;
1504 }
1505 
1506 static struct pnfs_layout_hdr *
1507 pnfs_find_alloc_layout(struct inode *ino,
1508 		       struct nfs_open_context *ctx,
1509 		       gfp_t gfp_flags)
1510 	__releases(&ino->i_lock)
1511 	__acquires(&ino->i_lock)
1512 {
1513 	struct nfs_inode *nfsi = NFS_I(ino);
1514 	struct pnfs_layout_hdr *new = NULL;
1515 
1516 	dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1517 
1518 	if (nfsi->layout != NULL)
1519 		goto out_existing;
1520 	spin_unlock(&ino->i_lock);
1521 	new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1522 	spin_lock(&ino->i_lock);
1523 
1524 	if (likely(nfsi->layout == NULL)) {	/* Won the race? */
1525 		nfsi->layout = new;
1526 		return new;
1527 	} else if (new != NULL)
1528 		pnfs_free_layout_hdr(new);
1529 out_existing:
1530 	pnfs_get_layout_hdr(nfsi->layout);
1531 	return nfsi->layout;
1532 }
1533 
1534 /*
1535  * iomode matching rules:
1536  * iomode	lseg	strict match
1537  *                      iomode
1538  * -----	-----	------ -----
1539  * ANY		READ	N/A    true
1540  * ANY		RW	N/A    true
1541  * RW		READ	N/A    false
1542  * RW		RW	N/A    true
1543  * READ		READ	N/A    true
1544  * READ		RW	true   false
1545  * READ		RW	false  true
1546  */
1547 static bool
1548 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1549 		 const struct pnfs_layout_range *range,
1550 		 bool strict_iomode)
1551 {
1552 	struct pnfs_layout_range range1;
1553 
1554 	if ((range->iomode == IOMODE_RW &&
1555 	     ls_range->iomode != IOMODE_RW) ||
1556 	    (range->iomode != ls_range->iomode &&
1557 	     strict_iomode == true) ||
1558 	    !pnfs_lseg_range_intersecting(ls_range, range))
1559 		return 0;
1560 
1561 	/* range1 covers only the first byte in the range */
1562 	range1 = *range;
1563 	range1.length = 1;
1564 	return pnfs_lseg_range_contained(ls_range, &range1);
1565 }
1566 
1567 /*
1568  * lookup range in layout
1569  */
1570 static struct pnfs_layout_segment *
1571 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1572 		struct pnfs_layout_range *range,
1573 		bool strict_iomode)
1574 {
1575 	struct pnfs_layout_segment *lseg, *ret = NULL;
1576 
1577 	dprintk("%s:Begin\n", __func__);
1578 
1579 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1580 		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1581 		    !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1582 		    pnfs_lseg_range_match(&lseg->pls_range, range,
1583 					  strict_iomode)) {
1584 			ret = pnfs_get_lseg(lseg);
1585 			break;
1586 		}
1587 	}
1588 
1589 	dprintk("%s:Return lseg %p ref %d\n",
1590 		__func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1591 	return ret;
1592 }
1593 
1594 /*
1595  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1596  * to the MDS or over pNFS
1597  *
1598  * The nfs_inode read_io and write_io fields are cumulative counters reset
1599  * when there are no layout segments. Note that in pnfs_update_layout iomode
1600  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1601  * WRITE request.
1602  *
1603  * A return of true means use MDS I/O.
1604  *
1605  * From rfc 5661:
1606  * If a file's size is smaller than the file size threshold, data accesses
1607  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1608  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1609  * server.  If both file size and I/O size are provided, the client SHOULD
1610  * reach or exceed  both thresholds before sending its read or write
1611  * requests to the data server.
1612  */
1613 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1614 				     struct inode *ino, int iomode)
1615 {
1616 	struct nfs4_threshold *t = ctx->mdsthreshold;
1617 	struct nfs_inode *nfsi = NFS_I(ino);
1618 	loff_t fsize = i_size_read(ino);
1619 	bool size = false, size_set = false, io = false, io_set = false, ret = false;
1620 
1621 	if (t == NULL)
1622 		return ret;
1623 
1624 	dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1625 		__func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1626 
1627 	switch (iomode) {
1628 	case IOMODE_READ:
1629 		if (t->bm & THRESHOLD_RD) {
1630 			dprintk("%s fsize %llu\n", __func__, fsize);
1631 			size_set = true;
1632 			if (fsize < t->rd_sz)
1633 				size = true;
1634 		}
1635 		if (t->bm & THRESHOLD_RD_IO) {
1636 			dprintk("%s nfsi->read_io %llu\n", __func__,
1637 				nfsi->read_io);
1638 			io_set = true;
1639 			if (nfsi->read_io < t->rd_io_sz)
1640 				io = true;
1641 		}
1642 		break;
1643 	case IOMODE_RW:
1644 		if (t->bm & THRESHOLD_WR) {
1645 			dprintk("%s fsize %llu\n", __func__, fsize);
1646 			size_set = true;
1647 			if (fsize < t->wr_sz)
1648 				size = true;
1649 		}
1650 		if (t->bm & THRESHOLD_WR_IO) {
1651 			dprintk("%s nfsi->write_io %llu\n", __func__,
1652 				nfsi->write_io);
1653 			io_set = true;
1654 			if (nfsi->write_io < t->wr_io_sz)
1655 				io = true;
1656 		}
1657 		break;
1658 	}
1659 	if (size_set && io_set) {
1660 		if (size && io)
1661 			ret = true;
1662 	} else if (size || io)
1663 		ret = true;
1664 
1665 	dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1666 	return ret;
1667 }
1668 
1669 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1670 {
1671 	/*
1672 	 * send layoutcommit as it can hold up layoutreturn due to lseg
1673 	 * reference
1674 	 */
1675 	pnfs_layoutcommit_inode(lo->plh_inode, false);
1676 	return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1677 				   nfs_wait_bit_killable,
1678 				   TASK_UNINTERRUPTIBLE);
1679 }
1680 
1681 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1682 {
1683 	unsigned long *bitlock = &lo->plh_flags;
1684 
1685 	clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1686 	smp_mb__after_atomic();
1687 	wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1688 }
1689 
1690 /*
1691  * Layout segment is retreived from the server if not cached.
1692  * The appropriate layout segment is referenced and returned to the caller.
1693  */
1694 struct pnfs_layout_segment *
1695 pnfs_update_layout(struct inode *ino,
1696 		   struct nfs_open_context *ctx,
1697 		   loff_t pos,
1698 		   u64 count,
1699 		   enum pnfs_iomode iomode,
1700 		   bool strict_iomode,
1701 		   gfp_t gfp_flags)
1702 {
1703 	struct pnfs_layout_range arg = {
1704 		.iomode = iomode,
1705 		.offset = pos,
1706 		.length = count,
1707 	};
1708 	unsigned pg_offset, seq;
1709 	struct nfs_server *server = NFS_SERVER(ino);
1710 	struct nfs_client *clp = server->nfs_client;
1711 	struct pnfs_layout_hdr *lo = NULL;
1712 	struct pnfs_layout_segment *lseg = NULL;
1713 	nfs4_stateid stateid;
1714 	long timeout = 0;
1715 	unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1716 	bool first;
1717 
1718 	if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1719 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1720 				 PNFS_UPDATE_LAYOUT_NO_PNFS);
1721 		goto out;
1722 	}
1723 
1724 	if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1725 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1726 				 PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1727 		goto out;
1728 	}
1729 
1730 	if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1731 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1732 				 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1733 		goto out;
1734 	}
1735 
1736 lookup_again:
1737 	nfs4_client_recover_expired_lease(clp);
1738 	first = false;
1739 	spin_lock(&ino->i_lock);
1740 	lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1741 	if (lo == NULL) {
1742 		spin_unlock(&ino->i_lock);
1743 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1744 				 PNFS_UPDATE_LAYOUT_NOMEM);
1745 		goto out;
1746 	}
1747 
1748 	/* Do we even need to bother with this? */
1749 	if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1750 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1751 				 PNFS_UPDATE_LAYOUT_BULK_RECALL);
1752 		dprintk("%s matches recall, use MDS\n", __func__);
1753 		goto out_unlock;
1754 	}
1755 
1756 	/* if LAYOUTGET already failed once we don't try again */
1757 	if (pnfs_layout_io_test_failed(lo, iomode)) {
1758 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1759 				 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1760 		goto out_unlock;
1761 	}
1762 
1763 	lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1764 	if (lseg) {
1765 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1766 				PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1767 		goto out_unlock;
1768 	}
1769 
1770 	if (!nfs4_valid_open_stateid(ctx->state)) {
1771 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1772 				PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1773 		goto out_unlock;
1774 	}
1775 
1776 	/*
1777 	 * Choose a stateid for the LAYOUTGET. If we don't have a layout
1778 	 * stateid, or it has been invalidated, then we must use the open
1779 	 * stateid.
1780 	 */
1781 	if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1782 
1783 		/*
1784 		 * The first layoutget for the file. Need to serialize per
1785 		 * RFC 5661 Errata 3208.
1786 		 */
1787 		if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1788 				     &lo->plh_flags)) {
1789 			spin_unlock(&ino->i_lock);
1790 			wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1791 				    TASK_UNINTERRUPTIBLE);
1792 			pnfs_put_layout_hdr(lo);
1793 			dprintk("%s retrying\n", __func__);
1794 			goto lookup_again;
1795 		}
1796 
1797 		first = true;
1798 		do {
1799 			seq = read_seqbegin(&ctx->state->seqlock);
1800 			nfs4_stateid_copy(&stateid, &ctx->state->stateid);
1801 		} while (read_seqretry(&ctx->state->seqlock, seq));
1802 	} else {
1803 		nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1804 	}
1805 
1806 	/*
1807 	 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1808 	 * for LAYOUTRETURN even if first is true.
1809 	 */
1810 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1811 		spin_unlock(&ino->i_lock);
1812 		dprintk("%s wait for layoutreturn\n", __func__);
1813 		if (pnfs_prepare_to_retry_layoutget(lo)) {
1814 			if (first)
1815 				pnfs_clear_first_layoutget(lo);
1816 			pnfs_put_layout_hdr(lo);
1817 			dprintk("%s retrying\n", __func__);
1818 			trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1819 					lseg, PNFS_UPDATE_LAYOUT_RETRY);
1820 			goto lookup_again;
1821 		}
1822 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1823 				PNFS_UPDATE_LAYOUT_RETURN);
1824 		goto out_put_layout_hdr;
1825 	}
1826 
1827 	if (pnfs_layoutgets_blocked(lo)) {
1828 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1829 				PNFS_UPDATE_LAYOUT_BLOCKED);
1830 		goto out_unlock;
1831 	}
1832 	atomic_inc(&lo->plh_outstanding);
1833 	spin_unlock(&ino->i_lock);
1834 
1835 	if (list_empty(&lo->plh_layouts)) {
1836 		/* The lo must be on the clp list if there is any
1837 		 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1838 		 */
1839 		spin_lock(&clp->cl_lock);
1840 		if (list_empty(&lo->plh_layouts))
1841 			list_add_tail(&lo->plh_layouts, &server->layouts);
1842 		spin_unlock(&clp->cl_lock);
1843 	}
1844 
1845 	pg_offset = arg.offset & ~PAGE_MASK;
1846 	if (pg_offset) {
1847 		arg.offset -= pg_offset;
1848 		arg.length += pg_offset;
1849 	}
1850 	if (arg.length != NFS4_MAX_UINT64)
1851 		arg.length = PAGE_ALIGN(arg.length);
1852 
1853 	lseg = send_layoutget(lo, ctx, &stateid, &arg, &timeout, gfp_flags);
1854 	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1855 				 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1856 	atomic_dec(&lo->plh_outstanding);
1857 	if (IS_ERR(lseg)) {
1858 		switch(PTR_ERR(lseg)) {
1859 		case -EBUSY:
1860 			if (time_after(jiffies, giveup))
1861 				lseg = NULL;
1862 			break;
1863 		case -ERECALLCONFLICT:
1864 			/* Huh? We hold no layouts, how is there a recall? */
1865 			if (first) {
1866 				lseg = NULL;
1867 				break;
1868 			}
1869 			/* Destroy the existing layout and start over */
1870 			if (time_after(jiffies, giveup))
1871 				pnfs_destroy_layout(NFS_I(ino));
1872 			/* Fallthrough */
1873 		case -EAGAIN:
1874 			break;
1875 		default:
1876 			if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1877 				pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1878 				lseg = NULL;
1879 			}
1880 			goto out_put_layout_hdr;
1881 		}
1882 		if (lseg) {
1883 			if (first)
1884 				pnfs_clear_first_layoutget(lo);
1885 			trace_pnfs_update_layout(ino, pos, count,
1886 				iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1887 			pnfs_put_layout_hdr(lo);
1888 			goto lookup_again;
1889 		}
1890 	} else {
1891 		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1892 	}
1893 
1894 out_put_layout_hdr:
1895 	if (first)
1896 		pnfs_clear_first_layoutget(lo);
1897 	pnfs_put_layout_hdr(lo);
1898 out:
1899 	dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1900 			"(%s, offset: %llu, length: %llu)\n",
1901 			__func__, ino->i_sb->s_id,
1902 			(unsigned long long)NFS_FILEID(ino),
1903 			IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1904 			iomode==IOMODE_RW ?  "read/write" : "read-only",
1905 			(unsigned long long)pos,
1906 			(unsigned long long)count);
1907 	return lseg;
1908 out_unlock:
1909 	spin_unlock(&ino->i_lock);
1910 	goto out_put_layout_hdr;
1911 }
1912 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1913 
1914 static bool
1915 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1916 {
1917 	switch (range->iomode) {
1918 	case IOMODE_READ:
1919 	case IOMODE_RW:
1920 		break;
1921 	default:
1922 		return false;
1923 	}
1924 	if (range->offset == NFS4_MAX_UINT64)
1925 		return false;
1926 	if (range->length == 0)
1927 		return false;
1928 	if (range->length != NFS4_MAX_UINT64 &&
1929 	    range->length > NFS4_MAX_UINT64 - range->offset)
1930 		return false;
1931 	return true;
1932 }
1933 
1934 struct pnfs_layout_segment *
1935 pnfs_layout_process(struct nfs4_layoutget *lgp)
1936 {
1937 	struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1938 	struct nfs4_layoutget_res *res = &lgp->res;
1939 	struct pnfs_layout_segment *lseg;
1940 	struct inode *ino = lo->plh_inode;
1941 	LIST_HEAD(free_me);
1942 
1943 	if (!pnfs_sanity_check_layout_range(&res->range))
1944 		return ERR_PTR(-EINVAL);
1945 
1946 	/* Inject layout blob into I/O device driver */
1947 	lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1948 	if (IS_ERR_OR_NULL(lseg)) {
1949 		if (!lseg)
1950 			lseg = ERR_PTR(-ENOMEM);
1951 
1952 		dprintk("%s: Could not allocate layout: error %ld\n",
1953 		       __func__, PTR_ERR(lseg));
1954 		return lseg;
1955 	}
1956 
1957 	pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
1958 
1959 	spin_lock(&ino->i_lock);
1960 	if (pnfs_layoutgets_blocked(lo)) {
1961 		dprintk("%s forget reply due to state\n", __func__);
1962 		goto out_forget;
1963 	}
1964 
1965 	if (!pnfs_layout_is_valid(lo)) {
1966 		/* We have a completely new layout */
1967 		pnfs_set_layout_stateid(lo, &res->stateid, true);
1968 	} else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1969 		/* existing state ID, make sure the sequence number matches. */
1970 		if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1971 			dprintk("%s forget reply due to sequence\n", __func__);
1972 			goto out_forget;
1973 		}
1974 		pnfs_set_layout_stateid(lo, &res->stateid, false);
1975 	} else {
1976 		/*
1977 		 * We got an entirely new state ID.  Mark all segments for the
1978 		 * inode invalid, and retry the layoutget
1979 		 */
1980 		pnfs_mark_layout_stateid_invalid(lo, &free_me);
1981 		goto out_forget;
1982 	}
1983 
1984 	pnfs_get_lseg(lseg);
1985 	pnfs_layout_insert_lseg(lo, lseg, &free_me);
1986 
1987 
1988 	if (res->return_on_close)
1989 		set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1990 
1991 	spin_unlock(&ino->i_lock);
1992 	pnfs_free_lseg_list(&free_me);
1993 	return lseg;
1994 
1995 out_forget:
1996 	spin_unlock(&ino->i_lock);
1997 	lseg->pls_layout = lo;
1998 	NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1999 	if (!pnfs_layout_is_valid(lo))
2000 		nfs_commit_inode(ino, 0);
2001 	return ERR_PTR(-EAGAIN);
2002 }
2003 
2004 /**
2005  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2006  * @lo: pointer to layout header
2007  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2008  * @return_range: describe layout segment ranges to be returned
2009  *
2010  * This function is mainly intended for use by layoutrecall. It attempts
2011  * to free the layout segment immediately, or else to mark it for return
2012  * as soon as its reference count drops to zero.
2013  */
2014 int
2015 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2016 				struct list_head *tmp_list,
2017 				const struct pnfs_layout_range *return_range,
2018 				u32 seq)
2019 {
2020 	struct pnfs_layout_segment *lseg, *next;
2021 	int remaining = 0;
2022 
2023 	dprintk("%s:Begin lo %p\n", __func__, lo);
2024 
2025 	if (list_empty(&lo->plh_segs))
2026 		return 0;
2027 
2028 	assert_spin_locked(&lo->plh_inode->i_lock);
2029 
2030 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2031 		if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2032 			dprintk("%s: marking lseg %p iomode %d "
2033 				"offset %llu length %llu\n", __func__,
2034 				lseg, lseg->pls_range.iomode,
2035 				lseg->pls_range.offset,
2036 				lseg->pls_range.length);
2037 			if (mark_lseg_invalid(lseg, tmp_list))
2038 				continue;
2039 			remaining++;
2040 			set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2041 		}
2042 
2043 	if (remaining)
2044 		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2045 
2046 	return remaining;
2047 }
2048 
2049 void pnfs_error_mark_layout_for_return(struct inode *inode,
2050 				       struct pnfs_layout_segment *lseg)
2051 {
2052 	struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2053 	struct pnfs_layout_range range = {
2054 		.iomode = lseg->pls_range.iomode,
2055 		.offset = 0,
2056 		.length = NFS4_MAX_UINT64,
2057 	};
2058 	bool return_now = false;
2059 
2060 	spin_lock(&inode->i_lock);
2061 	if (!pnfs_layout_is_valid(lo)) {
2062 		spin_unlock(&inode->i_lock);
2063 		return;
2064 	}
2065 	pnfs_set_plh_return_info(lo, range.iomode, 0);
2066 	/*
2067 	 * mark all matching lsegs so that we are sure to have no live
2068 	 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2069 	 * for how it works.
2070 	 */
2071 	if (!pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, &range, 0)) {
2072 		nfs4_stateid stateid;
2073 		enum pnfs_iomode iomode;
2074 
2075 		return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
2076 		spin_unlock(&inode->i_lock);
2077 		if (return_now)
2078 			pnfs_send_layoutreturn(lo, &stateid, iomode, false);
2079 	} else {
2080 		spin_unlock(&inode->i_lock);
2081 		nfs_commit_inode(inode, 0);
2082 	}
2083 }
2084 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2085 
2086 void
2087 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2088 {
2089 	if (pgio->pg_lseg == NULL ||
2090 	    test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2091 		return;
2092 	pnfs_put_lseg(pgio->pg_lseg);
2093 	pgio->pg_lseg = NULL;
2094 }
2095 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2096 
2097 /*
2098  * Check for any intersection between the request and the pgio->pg_lseg,
2099  * and if none, put this pgio->pg_lseg away.
2100  */
2101 static void
2102 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2103 {
2104 	if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2105 		pnfs_put_lseg(pgio->pg_lseg);
2106 		pgio->pg_lseg = NULL;
2107 	}
2108 }
2109 
2110 void
2111 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2112 {
2113 	u64 rd_size = req->wb_bytes;
2114 
2115 	pnfs_generic_pg_check_layout(pgio);
2116 	pnfs_generic_pg_check_range(pgio, req);
2117 	if (pgio->pg_lseg == NULL) {
2118 		if (pgio->pg_dreq == NULL)
2119 			rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2120 		else
2121 			rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2122 
2123 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2124 						   req->wb_context,
2125 						   req_offset(req),
2126 						   rd_size,
2127 						   IOMODE_READ,
2128 						   false,
2129 						   GFP_KERNEL);
2130 		if (IS_ERR(pgio->pg_lseg)) {
2131 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2132 			pgio->pg_lseg = NULL;
2133 			return;
2134 		}
2135 	}
2136 	/* If no lseg, fall back to read through mds */
2137 	if (pgio->pg_lseg == NULL)
2138 		nfs_pageio_reset_read_mds(pgio);
2139 
2140 }
2141 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2142 
2143 void
2144 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2145 			   struct nfs_page *req, u64 wb_size)
2146 {
2147 	pnfs_generic_pg_check_layout(pgio);
2148 	pnfs_generic_pg_check_range(pgio, req);
2149 	if (pgio->pg_lseg == NULL) {
2150 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2151 						   req->wb_context,
2152 						   req_offset(req),
2153 						   wb_size,
2154 						   IOMODE_RW,
2155 						   false,
2156 						   GFP_NOFS);
2157 		if (IS_ERR(pgio->pg_lseg)) {
2158 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2159 			pgio->pg_lseg = NULL;
2160 			return;
2161 		}
2162 	}
2163 	/* If no lseg, fall back to write through mds */
2164 	if (pgio->pg_lseg == NULL)
2165 		nfs_pageio_reset_write_mds(pgio);
2166 }
2167 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2168 
2169 void
2170 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2171 {
2172 	if (desc->pg_lseg) {
2173 		pnfs_put_lseg(desc->pg_lseg);
2174 		desc->pg_lseg = NULL;
2175 	}
2176 }
2177 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2178 
2179 /*
2180  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2181  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2182  */
2183 size_t
2184 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2185 		     struct nfs_page *prev, struct nfs_page *req)
2186 {
2187 	unsigned int size;
2188 	u64 seg_end, req_start, seg_left;
2189 
2190 	size = nfs_generic_pg_test(pgio, prev, req);
2191 	if (!size)
2192 		return 0;
2193 
2194 	/*
2195 	 * 'size' contains the number of bytes left in the current page (up
2196 	 * to the original size asked for in @req->wb_bytes).
2197 	 *
2198 	 * Calculate how many bytes are left in the layout segment
2199 	 * and if there are less bytes than 'size', return that instead.
2200 	 *
2201 	 * Please also note that 'end_offset' is actually the offset of the
2202 	 * first byte that lies outside the pnfs_layout_range. FIXME?
2203 	 *
2204 	 */
2205 	if (pgio->pg_lseg) {
2206 		seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2207 				     pgio->pg_lseg->pls_range.length);
2208 		req_start = req_offset(req);
2209 
2210 		/* start of request is past the last byte of this segment */
2211 		if (req_start >= seg_end)
2212 			return 0;
2213 
2214 		/* adjust 'size' iff there are fewer bytes left in the
2215 		 * segment than what nfs_generic_pg_test returned */
2216 		seg_left = seg_end - req_start;
2217 		if (seg_left < size)
2218 			size = (unsigned int)seg_left;
2219 	}
2220 
2221 	return size;
2222 }
2223 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2224 
2225 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2226 {
2227 	struct nfs_pageio_descriptor pgio;
2228 
2229 	/* Resend all requests through the MDS */
2230 	nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2231 			      hdr->completion_ops);
2232 	set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2233 	return nfs_pageio_resend(&pgio, hdr);
2234 }
2235 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2236 
2237 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2238 {
2239 
2240 	dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2241 	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2242 	    PNFS_LAYOUTRET_ON_ERROR) {
2243 		pnfs_return_layout(hdr->inode);
2244 	}
2245 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2246 		hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2247 }
2248 
2249 /*
2250  * Called by non rpc-based layout drivers
2251  */
2252 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2253 {
2254 	if (likely(!hdr->pnfs_error)) {
2255 		pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2256 				hdr->mds_offset + hdr->res.count);
2257 		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2258 	}
2259 	trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2260 	if (unlikely(hdr->pnfs_error))
2261 		pnfs_ld_handle_write_error(hdr);
2262 	hdr->mds_ops->rpc_release(hdr);
2263 }
2264 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2265 
2266 static void
2267 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2268 		struct nfs_pgio_header *hdr)
2269 {
2270 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2271 
2272 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2273 		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2274 		nfs_pageio_reset_write_mds(desc);
2275 		mirror->pg_recoalesce = 1;
2276 	}
2277 	nfs_pgio_data_destroy(hdr);
2278 	hdr->release(hdr);
2279 }
2280 
2281 static enum pnfs_try_status
2282 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2283 			const struct rpc_call_ops *call_ops,
2284 			struct pnfs_layout_segment *lseg,
2285 			int how)
2286 {
2287 	struct inode *inode = hdr->inode;
2288 	enum pnfs_try_status trypnfs;
2289 	struct nfs_server *nfss = NFS_SERVER(inode);
2290 
2291 	hdr->mds_ops = call_ops;
2292 
2293 	dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2294 		inode->i_ino, hdr->args.count, hdr->args.offset, how);
2295 	trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2296 	if (trypnfs != PNFS_NOT_ATTEMPTED)
2297 		nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2298 	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2299 	return trypnfs;
2300 }
2301 
2302 static void
2303 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2304 	      struct nfs_pgio_header *hdr, int how)
2305 {
2306 	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2307 	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2308 	enum pnfs_try_status trypnfs;
2309 
2310 	trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2311 	switch (trypnfs) {
2312 	case PNFS_NOT_ATTEMPTED:
2313 		pnfs_write_through_mds(desc, hdr);
2314 	case PNFS_ATTEMPTED:
2315 		break;
2316 	case PNFS_TRY_AGAIN:
2317 		/* cleanup hdr and prepare to redo pnfs */
2318 		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2319 			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2320 			list_splice_init(&hdr->pages, &mirror->pg_list);
2321 			mirror->pg_recoalesce = 1;
2322 		}
2323 		hdr->mds_ops->rpc_release(hdr);
2324 	}
2325 }
2326 
2327 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2328 {
2329 	pnfs_put_lseg(hdr->lseg);
2330 	nfs_pgio_header_free(hdr);
2331 }
2332 
2333 int
2334 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2335 {
2336 	struct nfs_pgio_header *hdr;
2337 	int ret;
2338 
2339 	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2340 	if (!hdr) {
2341 		desc->pg_error = -ENOMEM;
2342 		return desc->pg_error;
2343 	}
2344 	nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2345 
2346 	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2347 	ret = nfs_generic_pgio(desc, hdr);
2348 	if (!ret)
2349 		pnfs_do_write(desc, hdr, desc->pg_ioflags);
2350 
2351 	return ret;
2352 }
2353 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2354 
2355 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2356 {
2357 	struct nfs_pageio_descriptor pgio;
2358 
2359 	/* Resend all requests through the MDS */
2360 	nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2361 	return nfs_pageio_resend(&pgio, hdr);
2362 }
2363 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2364 
2365 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2366 {
2367 	dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2368 	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2369 	    PNFS_LAYOUTRET_ON_ERROR) {
2370 		pnfs_return_layout(hdr->inode);
2371 	}
2372 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2373 		hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2374 }
2375 
2376 /*
2377  * Called by non rpc-based layout drivers
2378  */
2379 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2380 {
2381 	if (likely(!hdr->pnfs_error))
2382 		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2383 	trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2384 	if (unlikely(hdr->pnfs_error))
2385 		pnfs_ld_handle_read_error(hdr);
2386 	hdr->mds_ops->rpc_release(hdr);
2387 }
2388 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2389 
2390 static void
2391 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2392 		struct nfs_pgio_header *hdr)
2393 {
2394 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2395 
2396 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2397 		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2398 		nfs_pageio_reset_read_mds(desc);
2399 		mirror->pg_recoalesce = 1;
2400 	}
2401 	nfs_pgio_data_destroy(hdr);
2402 	hdr->release(hdr);
2403 }
2404 
2405 /*
2406  * Call the appropriate parallel I/O subsystem read function.
2407  */
2408 static enum pnfs_try_status
2409 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2410 		       const struct rpc_call_ops *call_ops,
2411 		       struct pnfs_layout_segment *lseg)
2412 {
2413 	struct inode *inode = hdr->inode;
2414 	struct nfs_server *nfss = NFS_SERVER(inode);
2415 	enum pnfs_try_status trypnfs;
2416 
2417 	hdr->mds_ops = call_ops;
2418 
2419 	dprintk("%s: Reading ino:%lu %u@%llu\n",
2420 		__func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2421 
2422 	trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2423 	if (trypnfs != PNFS_NOT_ATTEMPTED)
2424 		nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2425 	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2426 	return trypnfs;
2427 }
2428 
2429 /* Resend all requests through pnfs. */
2430 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2431 {
2432 	struct nfs_pageio_descriptor pgio;
2433 
2434 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2435 		/* Prevent deadlocks with layoutreturn! */
2436 		pnfs_put_lseg(hdr->lseg);
2437 		hdr->lseg = NULL;
2438 
2439 		nfs_pageio_init_read(&pgio, hdr->inode, false,
2440 					hdr->completion_ops);
2441 		hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2442 	}
2443 }
2444 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2445 
2446 static void
2447 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2448 {
2449 	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2450 	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2451 	enum pnfs_try_status trypnfs;
2452 
2453 	trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2454 	switch (trypnfs) {
2455 	case PNFS_NOT_ATTEMPTED:
2456 		pnfs_read_through_mds(desc, hdr);
2457 	case PNFS_ATTEMPTED:
2458 		break;
2459 	case PNFS_TRY_AGAIN:
2460 		/* cleanup hdr and prepare to redo pnfs */
2461 		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2462 			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2463 			list_splice_init(&hdr->pages, &mirror->pg_list);
2464 			mirror->pg_recoalesce = 1;
2465 		}
2466 		hdr->mds_ops->rpc_release(hdr);
2467 	}
2468 }
2469 
2470 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2471 {
2472 	pnfs_put_lseg(hdr->lseg);
2473 	nfs_pgio_header_free(hdr);
2474 }
2475 
2476 int
2477 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2478 {
2479 	struct nfs_pgio_header *hdr;
2480 	int ret;
2481 
2482 	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2483 	if (!hdr) {
2484 		desc->pg_error = -ENOMEM;
2485 		return desc->pg_error;
2486 	}
2487 	nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2488 	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2489 	ret = nfs_generic_pgio(desc, hdr);
2490 	if (!ret)
2491 		pnfs_do_read(desc, hdr);
2492 	return ret;
2493 }
2494 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2495 
2496 static void pnfs_clear_layoutcommitting(struct inode *inode)
2497 {
2498 	unsigned long *bitlock = &NFS_I(inode)->flags;
2499 
2500 	clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2501 	smp_mb__after_atomic();
2502 	wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2503 }
2504 
2505 /*
2506  * There can be multiple RW segments.
2507  */
2508 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2509 {
2510 	struct pnfs_layout_segment *lseg;
2511 
2512 	list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2513 		if (lseg->pls_range.iomode == IOMODE_RW &&
2514 		    test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2515 			list_add(&lseg->pls_lc_list, listp);
2516 	}
2517 }
2518 
2519 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2520 {
2521 	struct pnfs_layout_segment *lseg, *tmp;
2522 
2523 	/* Matched by references in pnfs_set_layoutcommit */
2524 	list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2525 		list_del_init(&lseg->pls_lc_list);
2526 		pnfs_put_lseg(lseg);
2527 	}
2528 
2529 	pnfs_clear_layoutcommitting(inode);
2530 }
2531 
2532 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2533 {
2534 	pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2535 }
2536 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2537 
2538 void
2539 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2540 		loff_t end_pos)
2541 {
2542 	struct nfs_inode *nfsi = NFS_I(inode);
2543 	bool mark_as_dirty = false;
2544 
2545 	spin_lock(&inode->i_lock);
2546 	if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2547 		nfsi->layout->plh_lwb = end_pos;
2548 		mark_as_dirty = true;
2549 		dprintk("%s: Set layoutcommit for inode %lu ",
2550 			__func__, inode->i_ino);
2551 	} else if (end_pos > nfsi->layout->plh_lwb)
2552 		nfsi->layout->plh_lwb = end_pos;
2553 	if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2554 		/* references matched in nfs4_layoutcommit_release */
2555 		pnfs_get_lseg(lseg);
2556 	}
2557 	spin_unlock(&inode->i_lock);
2558 	dprintk("%s: lseg %p end_pos %llu\n",
2559 		__func__, lseg, nfsi->layout->plh_lwb);
2560 
2561 	/* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2562 	 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2563 	if (mark_as_dirty)
2564 		mark_inode_dirty_sync(inode);
2565 }
2566 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2567 
2568 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2569 {
2570 	struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2571 
2572 	if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2573 		nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2574 	pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2575 }
2576 
2577 /*
2578  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2579  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2580  * data to disk to allow the server to recover the data if it crashes.
2581  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2582  * is off, and a COMMIT is sent to a data server, or
2583  * if WRITEs to a data server return NFS_DATA_SYNC.
2584  */
2585 int
2586 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2587 {
2588 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2589 	struct nfs4_layoutcommit_data *data;
2590 	struct nfs_inode *nfsi = NFS_I(inode);
2591 	loff_t end_pos;
2592 	int status;
2593 
2594 	if (!pnfs_layoutcommit_outstanding(inode))
2595 		return 0;
2596 
2597 	dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2598 
2599 	status = -EAGAIN;
2600 	if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2601 		if (!sync)
2602 			goto out;
2603 		status = wait_on_bit_lock_action(&nfsi->flags,
2604 				NFS_INO_LAYOUTCOMMITTING,
2605 				nfs_wait_bit_killable,
2606 				TASK_KILLABLE);
2607 		if (status)
2608 			goto out;
2609 	}
2610 
2611 	status = -ENOMEM;
2612 	/* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2613 	data = kzalloc(sizeof(*data), GFP_NOFS);
2614 	if (!data)
2615 		goto clear_layoutcommitting;
2616 
2617 	status = 0;
2618 	spin_lock(&inode->i_lock);
2619 	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2620 		goto out_unlock;
2621 
2622 	INIT_LIST_HEAD(&data->lseg_list);
2623 	pnfs_list_write_lseg(inode, &data->lseg_list);
2624 
2625 	end_pos = nfsi->layout->plh_lwb;
2626 
2627 	nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2628 	spin_unlock(&inode->i_lock);
2629 
2630 	data->args.inode = inode;
2631 	data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2632 	nfs_fattr_init(&data->fattr);
2633 	data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2634 	data->res.fattr = &data->fattr;
2635 	if (end_pos != 0)
2636 		data->args.lastbytewritten = end_pos - 1;
2637 	else
2638 		data->args.lastbytewritten = U64_MAX;
2639 	data->res.server = NFS_SERVER(inode);
2640 
2641 	if (ld->prepare_layoutcommit) {
2642 		status = ld->prepare_layoutcommit(&data->args);
2643 		if (status) {
2644 			put_rpccred(data->cred);
2645 			spin_lock(&inode->i_lock);
2646 			set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2647 			if (end_pos > nfsi->layout->plh_lwb)
2648 				nfsi->layout->plh_lwb = end_pos;
2649 			goto out_unlock;
2650 		}
2651 	}
2652 
2653 
2654 	status = nfs4_proc_layoutcommit(data, sync);
2655 out:
2656 	if (status)
2657 		mark_inode_dirty_sync(inode);
2658 	dprintk("<-- %s status %d\n", __func__, status);
2659 	return status;
2660 out_unlock:
2661 	spin_unlock(&inode->i_lock);
2662 	kfree(data);
2663 clear_layoutcommitting:
2664 	pnfs_clear_layoutcommitting(inode);
2665 	goto out;
2666 }
2667 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2668 
2669 int
2670 pnfs_generic_sync(struct inode *inode, bool datasync)
2671 {
2672 	return pnfs_layoutcommit_inode(inode, true);
2673 }
2674 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2675 
2676 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2677 {
2678 	struct nfs4_threshold *thp;
2679 
2680 	thp = kzalloc(sizeof(*thp), GFP_NOFS);
2681 	if (!thp) {
2682 		dprintk("%s mdsthreshold allocation failed\n", __func__);
2683 		return NULL;
2684 	}
2685 	return thp;
2686 }
2687 
2688 #if IS_ENABLED(CONFIG_NFS_V4_2)
2689 int
2690 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2691 {
2692 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2693 	struct nfs_server *server = NFS_SERVER(inode);
2694 	struct nfs_inode *nfsi = NFS_I(inode);
2695 	struct nfs42_layoutstat_data *data;
2696 	struct pnfs_layout_hdr *hdr;
2697 	int status = 0;
2698 
2699 	if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2700 		goto out;
2701 
2702 	if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2703 		goto out;
2704 
2705 	if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2706 		goto out;
2707 
2708 	spin_lock(&inode->i_lock);
2709 	if (!NFS_I(inode)->layout) {
2710 		spin_unlock(&inode->i_lock);
2711 		goto out_clear_layoutstats;
2712 	}
2713 	hdr = NFS_I(inode)->layout;
2714 	pnfs_get_layout_hdr(hdr);
2715 	spin_unlock(&inode->i_lock);
2716 
2717 	data = kzalloc(sizeof(*data), gfp_flags);
2718 	if (!data) {
2719 		status = -ENOMEM;
2720 		goto out_put;
2721 	}
2722 
2723 	data->args.fh = NFS_FH(inode);
2724 	data->args.inode = inode;
2725 	status = ld->prepare_layoutstats(&data->args);
2726 	if (status)
2727 		goto out_free;
2728 
2729 	status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2730 
2731 out:
2732 	dprintk("%s returns %d\n", __func__, status);
2733 	return status;
2734 
2735 out_free:
2736 	kfree(data);
2737 out_put:
2738 	pnfs_put_layout_hdr(hdr);
2739 out_clear_layoutstats:
2740 	smp_mb__before_atomic();
2741 	clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2742 	smp_mb__after_atomic();
2743 	goto out;
2744 }
2745 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2746 #endif
2747 
2748 unsigned int layoutstats_timer;
2749 module_param(layoutstats_timer, uint, 0644);
2750 EXPORT_SYMBOL_GPL(layoutstats_timer);
2751