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 "internal.h"
34#include "pnfs.h"
35#include "iostat.h"
36#include "nfs4trace.h"
37#include "delegation.h"
38#include "nfs42.h"
39
40#define NFSDBG_FACILITY		NFSDBG_PNFS
41#define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
42
43/* Locking:
44 *
45 * pnfs_spinlock:
46 *      protects pnfs_modules_tbl.
47 */
48static DEFINE_SPINLOCK(pnfs_spinlock);
49
50/*
51 * pnfs_modules_tbl holds all pnfs modules
52 */
53static LIST_HEAD(pnfs_modules_tbl);
54
55static int
56pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
57		       enum pnfs_iomode iomode, bool sync);
58
59/* Return the registered pnfs layout driver module matching given id */
60static struct pnfs_layoutdriver_type *
61find_pnfs_driver_locked(u32 id)
62{
63	struct pnfs_layoutdriver_type *local;
64
65	list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
66		if (local->id == id)
67			goto out;
68	local = NULL;
69out:
70	dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
71	return local;
72}
73
74static struct pnfs_layoutdriver_type *
75find_pnfs_driver(u32 id)
76{
77	struct pnfs_layoutdriver_type *local;
78
79	spin_lock(&pnfs_spinlock);
80	local = find_pnfs_driver_locked(id);
81	if (local != NULL && !try_module_get(local->owner)) {
82		dprintk("%s: Could not grab reference on module\n", __func__);
83		local = NULL;
84	}
85	spin_unlock(&pnfs_spinlock);
86	return local;
87}
88
89void
90unset_pnfs_layoutdriver(struct nfs_server *nfss)
91{
92	if (nfss->pnfs_curr_ld) {
93		if (nfss->pnfs_curr_ld->clear_layoutdriver)
94			nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
95		/* Decrement the MDS count. Purge the deviceid cache if zero */
96		if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
97			nfs4_deviceid_purge_client(nfss->nfs_client);
98		module_put(nfss->pnfs_curr_ld->owner);
99	}
100	nfss->pnfs_curr_ld = NULL;
101}
102
103/*
104 * Try to set the server's pnfs module to the pnfs layout type specified by id.
105 * Currently only one pNFS layout driver per filesystem is supported.
106 *
107 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
108 */
109void
110set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
111		      u32 id)
112{
113	struct pnfs_layoutdriver_type *ld_type = NULL;
114
115	if (id == 0)
116		goto out_no_driver;
117	if (!(server->nfs_client->cl_exchange_flags &
118		 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
119		printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
120			__func__, id, server->nfs_client->cl_exchange_flags);
121		goto out_no_driver;
122	}
123	ld_type = find_pnfs_driver(id);
124	if (!ld_type) {
125		request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
126		ld_type = find_pnfs_driver(id);
127		if (!ld_type) {
128			dprintk("%s: No pNFS module found for %u.\n",
129				__func__, id);
130			goto out_no_driver;
131		}
132	}
133	server->pnfs_curr_ld = ld_type;
134	if (ld_type->set_layoutdriver
135	    && ld_type->set_layoutdriver(server, mntfh)) {
136		printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
137			"driver %u.\n", __func__, id);
138		module_put(ld_type->owner);
139		goto out_no_driver;
140	}
141	/* Bump the MDS count */
142	atomic_inc(&server->nfs_client->cl_mds_count);
143
144	dprintk("%s: pNFS module for %u set\n", __func__, id);
145	return;
146
147out_no_driver:
148	dprintk("%s: Using NFSv4 I/O\n", __func__);
149	server->pnfs_curr_ld = NULL;
150}
151
152int
153pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
154{
155	int status = -EINVAL;
156	struct pnfs_layoutdriver_type *tmp;
157
158	if (ld_type->id == 0) {
159		printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
160		return status;
161	}
162	if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
163		printk(KERN_ERR "NFS: %s Layout driver must provide "
164		       "alloc_lseg and free_lseg.\n", __func__);
165		return status;
166	}
167
168	spin_lock(&pnfs_spinlock);
169	tmp = find_pnfs_driver_locked(ld_type->id);
170	if (!tmp) {
171		list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
172		status = 0;
173		dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
174			ld_type->name);
175	} else {
176		printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
177			__func__, ld_type->id);
178	}
179	spin_unlock(&pnfs_spinlock);
180
181	return status;
182}
183EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
184
185void
186pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
187{
188	dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
189	spin_lock(&pnfs_spinlock);
190	list_del(&ld_type->pnfs_tblid);
191	spin_unlock(&pnfs_spinlock);
192}
193EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
194
195/*
196 * pNFS client layout cache
197 */
198
199/* Need to hold i_lock if caller does not already hold reference */
200void
201pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
202{
203	atomic_inc(&lo->plh_refcount);
204}
205
206static struct pnfs_layout_hdr *
207pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
208{
209	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
210	return ld->alloc_layout_hdr(ino, gfp_flags);
211}
212
213static void
214pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
215{
216	struct nfs_server *server = NFS_SERVER(lo->plh_inode);
217	struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
218
219	if (!list_empty(&lo->plh_layouts)) {
220		struct nfs_client *clp = server->nfs_client;
221
222		spin_lock(&clp->cl_lock);
223		list_del_init(&lo->plh_layouts);
224		spin_unlock(&clp->cl_lock);
225	}
226	put_rpccred(lo->plh_lc_cred);
227	return ld->free_layout_hdr(lo);
228}
229
230static void
231pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
232{
233	struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
234	dprintk("%s: freeing layout cache %p\n", __func__, lo);
235	nfsi->layout = NULL;
236	/* Reset MDS Threshold I/O counters */
237	nfsi->write_io = 0;
238	nfsi->read_io = 0;
239}
240
241void
242pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
243{
244	struct inode *inode = lo->plh_inode;
245
246	if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
247		if (!list_empty(&lo->plh_segs))
248			WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
249		pnfs_detach_layout_hdr(lo);
250		spin_unlock(&inode->i_lock);
251		pnfs_free_layout_hdr(lo);
252	}
253}
254
255static int
256pnfs_iomode_to_fail_bit(u32 iomode)
257{
258	return iomode == IOMODE_RW ?
259		NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
260}
261
262static void
263pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
264{
265	lo->plh_retry_timestamp = jiffies;
266	if (!test_and_set_bit(fail_bit, &lo->plh_flags))
267		atomic_inc(&lo->plh_refcount);
268}
269
270static void
271pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
272{
273	if (test_and_clear_bit(fail_bit, &lo->plh_flags))
274		atomic_dec(&lo->plh_refcount);
275}
276
277static void
278pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
279{
280	struct inode *inode = lo->plh_inode;
281	struct pnfs_layout_range range = {
282		.iomode = iomode,
283		.offset = 0,
284		.length = NFS4_MAX_UINT64,
285	};
286	LIST_HEAD(head);
287
288	spin_lock(&inode->i_lock);
289	pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
290	pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
291	spin_unlock(&inode->i_lock);
292	pnfs_free_lseg_list(&head);
293	dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
294			iomode == IOMODE_RW ?  "RW" : "READ");
295}
296
297static bool
298pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
299{
300	unsigned long start, end;
301	int fail_bit = pnfs_iomode_to_fail_bit(iomode);
302
303	if (test_bit(fail_bit, &lo->plh_flags) == 0)
304		return false;
305	end = jiffies;
306	start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
307	if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
308		/* It is time to retry the failed layoutgets */
309		pnfs_layout_clear_fail_bit(lo, fail_bit);
310		return false;
311	}
312	return true;
313}
314
315static void
316init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
317{
318	INIT_LIST_HEAD(&lseg->pls_list);
319	INIT_LIST_HEAD(&lseg->pls_lc_list);
320	atomic_set(&lseg->pls_refcount, 1);
321	smp_mb();
322	set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
323	lseg->pls_layout = lo;
324}
325
326static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
327{
328	struct inode *ino = lseg->pls_layout->plh_inode;
329
330	NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
331}
332
333static void
334pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
335		struct pnfs_layout_segment *lseg)
336{
337	struct inode *inode = lo->plh_inode;
338
339	WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
340	list_del_init(&lseg->pls_list);
341	/* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
342	atomic_dec(&lo->plh_refcount);
343	if (list_empty(&lo->plh_segs))
344		clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
345	rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
346}
347
348/* Return true if layoutreturn is needed */
349static bool
350pnfs_layout_need_return(struct pnfs_layout_hdr *lo,
351			struct pnfs_layout_segment *lseg)
352{
353	struct pnfs_layout_segment *s;
354
355	if (!test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
356		return false;
357
358	list_for_each_entry(s, &lo->plh_segs, pls_list)
359		if (s != lseg && test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
360			return false;
361
362	return true;
363}
364
365static bool
366pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
367{
368	if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
369		return false;
370	lo->plh_return_iomode = 0;
371	pnfs_get_layout_hdr(lo);
372	clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, &lo->plh_flags);
373	return true;
374}
375
376static void pnfs_layoutreturn_before_put_lseg(struct pnfs_layout_segment *lseg,
377		struct pnfs_layout_hdr *lo, struct inode *inode)
378{
379	lo = lseg->pls_layout;
380	inode = lo->plh_inode;
381
382	spin_lock(&inode->i_lock);
383	if (pnfs_layout_need_return(lo, lseg)) {
384		nfs4_stateid stateid;
385		enum pnfs_iomode iomode;
386		bool send;
387
388		stateid = lo->plh_stateid;
389		iomode = lo->plh_return_iomode;
390		send = pnfs_prepare_layoutreturn(lo);
391		spin_unlock(&inode->i_lock);
392		if (send) {
393			/* Send an async layoutreturn so we dont deadlock */
394			pnfs_send_layoutreturn(lo, stateid, iomode, false);
395		}
396	} else
397		spin_unlock(&inode->i_lock);
398}
399
400void
401pnfs_put_lseg(struct pnfs_layout_segment *lseg)
402{
403	struct pnfs_layout_hdr *lo;
404	struct inode *inode;
405
406	if (!lseg)
407		return;
408
409	dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
410		atomic_read(&lseg->pls_refcount),
411		test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
412
413	/* Handle the case where refcount != 1 */
414	if (atomic_add_unless(&lseg->pls_refcount, -1, 1))
415		return;
416
417	lo = lseg->pls_layout;
418	inode = lo->plh_inode;
419	/* Do we need a layoutreturn? */
420	if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
421		pnfs_layoutreturn_before_put_lseg(lseg, lo, inode);
422
423	if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
424		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
425			spin_unlock(&inode->i_lock);
426			return;
427		}
428		pnfs_get_layout_hdr(lo);
429		pnfs_layout_remove_lseg(lo, lseg);
430		spin_unlock(&inode->i_lock);
431		pnfs_free_lseg(lseg);
432		pnfs_put_layout_hdr(lo);
433	}
434}
435EXPORT_SYMBOL_GPL(pnfs_put_lseg);
436
437static void pnfs_free_lseg_async_work(struct work_struct *work)
438{
439	struct pnfs_layout_segment *lseg;
440	struct pnfs_layout_hdr *lo;
441
442	lseg = container_of(work, struct pnfs_layout_segment, pls_work);
443	lo = lseg->pls_layout;
444
445	pnfs_free_lseg(lseg);
446	pnfs_put_layout_hdr(lo);
447}
448
449static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
450{
451	INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
452	schedule_work(&lseg->pls_work);
453}
454
455void
456pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
457{
458	if (!lseg)
459		return;
460
461	assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
462
463	dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
464		atomic_read(&lseg->pls_refcount),
465		test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
466	if (atomic_dec_and_test(&lseg->pls_refcount)) {
467		struct pnfs_layout_hdr *lo = lseg->pls_layout;
468		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
469			return;
470		pnfs_get_layout_hdr(lo);
471		pnfs_layout_remove_lseg(lo, lseg);
472		pnfs_free_lseg_async(lseg);
473	}
474}
475EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
476
477static u64
478end_offset(u64 start, u64 len)
479{
480	u64 end;
481
482	end = start + len;
483	return end >= start ? end : NFS4_MAX_UINT64;
484}
485
486/*
487 * is l2 fully contained in l1?
488 *   start1                             end1
489 *   [----------------------------------)
490 *           start2           end2
491 *           [----------------)
492 */
493static bool
494pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
495		 const struct pnfs_layout_range *l2)
496{
497	u64 start1 = l1->offset;
498	u64 end1 = end_offset(start1, l1->length);
499	u64 start2 = l2->offset;
500	u64 end2 = end_offset(start2, l2->length);
501
502	return (start1 <= start2) && (end1 >= end2);
503}
504
505/*
506 * is l1 and l2 intersecting?
507 *   start1                             end1
508 *   [----------------------------------)
509 *                              start2           end2
510 *                              [----------------)
511 */
512static bool
513pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
514		    const struct pnfs_layout_range *l2)
515{
516	u64 start1 = l1->offset;
517	u64 end1 = end_offset(start1, l1->length);
518	u64 start2 = l2->offset;
519	u64 end2 = end_offset(start2, l2->length);
520
521	return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
522	       (end2 == NFS4_MAX_UINT64 || end2 > start1);
523}
524
525static bool
526should_free_lseg(const struct pnfs_layout_range *lseg_range,
527		 const struct pnfs_layout_range *recall_range)
528{
529	return (recall_range->iomode == IOMODE_ANY ||
530		lseg_range->iomode == recall_range->iomode) &&
531	       pnfs_lseg_range_intersecting(lseg_range, recall_range);
532}
533
534static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
535		struct list_head *tmp_list)
536{
537	if (!atomic_dec_and_test(&lseg->pls_refcount))
538		return false;
539	pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
540	list_add(&lseg->pls_list, tmp_list);
541	return true;
542}
543
544/* Returns 1 if lseg is removed from list, 0 otherwise */
545static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
546			     struct list_head *tmp_list)
547{
548	int rv = 0;
549
550	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
551		/* Remove the reference keeping the lseg in the
552		 * list.  It will now be removed when all
553		 * outstanding io is finished.
554		 */
555		dprintk("%s: lseg %p ref %d\n", __func__, lseg,
556			atomic_read(&lseg->pls_refcount));
557		if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
558			rv = 1;
559	}
560	return rv;
561}
562
563/* Returns count of number of matching invalid lsegs remaining in list
564 * after call.
565 */
566int
567pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
568			    struct list_head *tmp_list,
569			    struct pnfs_layout_range *recall_range)
570{
571	struct pnfs_layout_segment *lseg, *next;
572	int invalid = 0, removed = 0;
573
574	dprintk("%s:Begin lo %p\n", __func__, lo);
575
576	if (list_empty(&lo->plh_segs))
577		return 0;
578	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
579		if (!recall_range ||
580		    should_free_lseg(&lseg->pls_range, recall_range)) {
581			dprintk("%s: freeing lseg %p iomode %d "
582				"offset %llu length %llu\n", __func__,
583				lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
584				lseg->pls_range.length);
585			invalid++;
586			removed += mark_lseg_invalid(lseg, tmp_list);
587		}
588	dprintk("%s:Return %i\n", __func__, invalid - removed);
589	return invalid - removed;
590}
591
592/* note free_me must contain lsegs from a single layout_hdr */
593void
594pnfs_free_lseg_list(struct list_head *free_me)
595{
596	struct pnfs_layout_segment *lseg, *tmp;
597
598	if (list_empty(free_me))
599		return;
600
601	list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
602		list_del(&lseg->pls_list);
603		pnfs_free_lseg(lseg);
604	}
605}
606
607void
608pnfs_destroy_layout(struct nfs_inode *nfsi)
609{
610	struct pnfs_layout_hdr *lo;
611	LIST_HEAD(tmp_list);
612
613	spin_lock(&nfsi->vfs_inode.i_lock);
614	lo = nfsi->layout;
615	if (lo) {
616		lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
617		pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
618		pnfs_get_layout_hdr(lo);
619		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
620		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
621		pnfs_clear_retry_layoutget(lo);
622		spin_unlock(&nfsi->vfs_inode.i_lock);
623		pnfs_free_lseg_list(&tmp_list);
624		pnfs_put_layout_hdr(lo);
625	} else
626		spin_unlock(&nfsi->vfs_inode.i_lock);
627}
628EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
629
630static bool
631pnfs_layout_add_bulk_destroy_list(struct inode *inode,
632		struct list_head *layout_list)
633{
634	struct pnfs_layout_hdr *lo;
635	bool ret = false;
636
637	spin_lock(&inode->i_lock);
638	lo = NFS_I(inode)->layout;
639	if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
640		pnfs_get_layout_hdr(lo);
641		list_add(&lo->plh_bulk_destroy, layout_list);
642		ret = true;
643	}
644	spin_unlock(&inode->i_lock);
645	return ret;
646}
647
648/* Caller must hold rcu_read_lock and clp->cl_lock */
649static int
650pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
651		struct nfs_server *server,
652		struct list_head *layout_list)
653{
654	struct pnfs_layout_hdr *lo, *next;
655	struct inode *inode;
656
657	list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
658		inode = igrab(lo->plh_inode);
659		if (inode == NULL)
660			continue;
661		list_del_init(&lo->plh_layouts);
662		if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
663			continue;
664		rcu_read_unlock();
665		spin_unlock(&clp->cl_lock);
666		iput(inode);
667		spin_lock(&clp->cl_lock);
668		rcu_read_lock();
669		return -EAGAIN;
670	}
671	return 0;
672}
673
674static int
675pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
676		bool is_bulk_recall)
677{
678	struct pnfs_layout_hdr *lo;
679	struct inode *inode;
680	struct pnfs_layout_range range = {
681		.iomode = IOMODE_ANY,
682		.offset = 0,
683		.length = NFS4_MAX_UINT64,
684	};
685	LIST_HEAD(lseg_list);
686	int ret = 0;
687
688	while (!list_empty(layout_list)) {
689		lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
690				plh_bulk_destroy);
691		dprintk("%s freeing layout for inode %lu\n", __func__,
692			lo->plh_inode->i_ino);
693		inode = lo->plh_inode;
694
695		pnfs_layoutcommit_inode(inode, false);
696
697		spin_lock(&inode->i_lock);
698		list_del_init(&lo->plh_bulk_destroy);
699		lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
700		if (is_bulk_recall)
701			set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
702		if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
703			ret = -EAGAIN;
704		spin_unlock(&inode->i_lock);
705		pnfs_free_lseg_list(&lseg_list);
706		pnfs_put_layout_hdr(lo);
707		iput(inode);
708	}
709	return ret;
710}
711
712int
713pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
714		struct nfs_fsid *fsid,
715		bool is_recall)
716{
717	struct nfs_server *server;
718	LIST_HEAD(layout_list);
719
720	spin_lock(&clp->cl_lock);
721	rcu_read_lock();
722restart:
723	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
724		if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
725			continue;
726		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
727				server,
728				&layout_list) != 0)
729			goto restart;
730	}
731	rcu_read_unlock();
732	spin_unlock(&clp->cl_lock);
733
734	if (list_empty(&layout_list))
735		return 0;
736	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
737}
738
739int
740pnfs_destroy_layouts_byclid(struct nfs_client *clp,
741		bool is_recall)
742{
743	struct nfs_server *server;
744	LIST_HEAD(layout_list);
745
746	spin_lock(&clp->cl_lock);
747	rcu_read_lock();
748restart:
749	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
750		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
751					server,
752					&layout_list) != 0)
753			goto restart;
754	}
755	rcu_read_unlock();
756	spin_unlock(&clp->cl_lock);
757
758	if (list_empty(&layout_list))
759		return 0;
760	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
761}
762
763/*
764 * Called by the state manger to remove all layouts established under an
765 * expired lease.
766 */
767void
768pnfs_destroy_all_layouts(struct nfs_client *clp)
769{
770	nfs4_deviceid_mark_client_invalid(clp);
771	nfs4_deviceid_purge_client(clp);
772
773	pnfs_destroy_layouts_byclid(clp, false);
774}
775
776/*
777 * Compare 2 layout stateid sequence ids, to see which is newer,
778 * taking into account wraparound issues.
779 */
780static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
781{
782	return (s32)(s1 - s2) > 0;
783}
784
785/* update lo->plh_stateid with new if is more recent */
786void
787pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
788			bool update_barrier)
789{
790	u32 oldseq, newseq, new_barrier;
791	int empty = list_empty(&lo->plh_segs);
792
793	oldseq = be32_to_cpu(lo->plh_stateid.seqid);
794	newseq = be32_to_cpu(new->seqid);
795	if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
796		nfs4_stateid_copy(&lo->plh_stateid, new);
797		if (update_barrier) {
798			new_barrier = be32_to_cpu(new->seqid);
799		} else {
800			/* Because of wraparound, we want to keep the barrier
801			 * "close" to the current seqids.
802			 */
803			new_barrier = newseq - atomic_read(&lo->plh_outstanding);
804		}
805		if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
806			lo->plh_barrier = new_barrier;
807	}
808}
809
810static bool
811pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
812		const nfs4_stateid *stateid)
813{
814	u32 seqid = be32_to_cpu(stateid->seqid);
815
816	return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
817}
818
819/* lget is set to 1 if called from inside send_layoutget call chain */
820static bool
821pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
822{
823	return lo->plh_block_lgets ||
824		test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
825}
826
827int
828pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
829			      struct pnfs_layout_range *range,
830			      struct nfs4_state *open_state)
831{
832	int status = 0;
833
834	dprintk("--> %s\n", __func__);
835	spin_lock(&lo->plh_inode->i_lock);
836	if (pnfs_layoutgets_blocked(lo)) {
837		status = -EAGAIN;
838	} else if (!nfs4_valid_open_stateid(open_state)) {
839		status = -EBADF;
840	} else if (list_empty(&lo->plh_segs) ||
841		   test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
842		int seq;
843
844		do {
845			seq = read_seqbegin(&open_state->seqlock);
846			nfs4_stateid_copy(dst, &open_state->stateid);
847		} while (read_seqretry(&open_state->seqlock, seq));
848	} else
849		nfs4_stateid_copy(dst, &lo->plh_stateid);
850	spin_unlock(&lo->plh_inode->i_lock);
851	dprintk("<-- %s\n", __func__);
852	return status;
853}
854
855/*
856* Get layout from server.
857*    for now, assume that whole file layouts are requested.
858*    arg->offset: 0
859*    arg->length: all ones
860*/
861static struct pnfs_layout_segment *
862send_layoutget(struct pnfs_layout_hdr *lo,
863	   struct nfs_open_context *ctx,
864	   struct pnfs_layout_range *range,
865	   gfp_t gfp_flags)
866{
867	struct inode *ino = lo->plh_inode;
868	struct nfs_server *server = NFS_SERVER(ino);
869	struct nfs4_layoutget *lgp;
870	struct pnfs_layout_segment *lseg;
871	loff_t i_size;
872
873	dprintk("--> %s\n", __func__);
874
875	/*
876	 * Synchronously retrieve layout information from server and
877	 * store in lseg. If we race with a concurrent seqid morphing
878	 * op, then re-send the LAYOUTGET.
879	 */
880	do {
881		lgp = kzalloc(sizeof(*lgp), gfp_flags);
882		if (lgp == NULL)
883			return NULL;
884
885		i_size = i_size_read(ino);
886
887		lgp->args.minlength = PAGE_CACHE_SIZE;
888		if (lgp->args.minlength > range->length)
889			lgp->args.minlength = range->length;
890		if (range->iomode == IOMODE_READ) {
891			if (range->offset >= i_size)
892				lgp->args.minlength = 0;
893			else if (i_size - range->offset < lgp->args.minlength)
894				lgp->args.minlength = i_size - range->offset;
895		}
896		lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
897		lgp->args.range = *range;
898		lgp->args.type = server->pnfs_curr_ld->id;
899		lgp->args.inode = ino;
900		lgp->args.ctx = get_nfs_open_context(ctx);
901		lgp->gfp_flags = gfp_flags;
902		lgp->cred = lo->plh_lc_cred;
903
904		lseg = nfs4_proc_layoutget(lgp, gfp_flags);
905	} while (lseg == ERR_PTR(-EAGAIN));
906
907	if (IS_ERR(lseg)) {
908		switch (PTR_ERR(lseg)) {
909		case -ENOMEM:
910		case -ERESTARTSYS:
911			break;
912		default:
913			/* remember that LAYOUTGET failed and suspend trying */
914			pnfs_layout_io_set_failed(lo, range->iomode);
915		}
916		return NULL;
917	} else
918		pnfs_layout_clear_fail_bit(lo,
919				pnfs_iomode_to_fail_bit(range->iomode));
920
921	return lseg;
922}
923
924static void pnfs_clear_layoutcommit(struct inode *inode,
925		struct list_head *head)
926{
927	struct nfs_inode *nfsi = NFS_I(inode);
928	struct pnfs_layout_segment *lseg, *tmp;
929
930	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
931		return;
932	list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
933		if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
934			continue;
935		pnfs_lseg_dec_and_remove_zero(lseg, head);
936	}
937}
938
939void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
940{
941	clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
942	smp_mb__after_atomic();
943	wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
944	rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
945}
946
947static int
948pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
949		       enum pnfs_iomode iomode, bool sync)
950{
951	struct inode *ino = lo->plh_inode;
952	struct nfs4_layoutreturn *lrp;
953	int status = 0;
954
955	lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
956	if (unlikely(lrp == NULL)) {
957		status = -ENOMEM;
958		spin_lock(&ino->i_lock);
959		pnfs_clear_layoutreturn_waitbit(lo);
960		spin_unlock(&ino->i_lock);
961		pnfs_put_layout_hdr(lo);
962		goto out;
963	}
964
965	lrp->args.stateid = stateid;
966	lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
967	lrp->args.inode = ino;
968	lrp->args.range.iomode = iomode;
969	lrp->args.range.offset = 0;
970	lrp->args.range.length = NFS4_MAX_UINT64;
971	lrp->args.layout = lo;
972	lrp->clp = NFS_SERVER(ino)->nfs_client;
973	lrp->cred = lo->plh_lc_cred;
974
975	status = nfs4_proc_layoutreturn(lrp, sync);
976out:
977	dprintk("<-- %s status: %d\n", __func__, status);
978	return status;
979}
980
981/*
982 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
983 * when the layout segment list is empty.
984 *
985 * Note that a pnfs_layout_hdr can exist with an empty layout segment
986 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
987 * deviceid is marked invalid.
988 */
989int
990_pnfs_return_layout(struct inode *ino)
991{
992	struct pnfs_layout_hdr *lo = NULL;
993	struct nfs_inode *nfsi = NFS_I(ino);
994	LIST_HEAD(tmp_list);
995	nfs4_stateid stateid;
996	int status = 0, empty;
997	bool send;
998
999	dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1000
1001	spin_lock(&ino->i_lock);
1002	lo = nfsi->layout;
1003	if (!lo) {
1004		spin_unlock(&ino->i_lock);
1005		dprintk("NFS: %s no layout to return\n", __func__);
1006		goto out;
1007	}
1008	stateid = nfsi->layout->plh_stateid;
1009	/* Reference matched in nfs4_layoutreturn_release */
1010	pnfs_get_layout_hdr(lo);
1011	empty = list_empty(&lo->plh_segs);
1012	pnfs_clear_layoutcommit(ino, &tmp_list);
1013	pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
1014
1015	if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1016		struct pnfs_layout_range range = {
1017			.iomode		= IOMODE_ANY,
1018			.offset		= 0,
1019			.length		= NFS4_MAX_UINT64,
1020		};
1021		NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1022	}
1023
1024	/* Don't send a LAYOUTRETURN if list was initially empty */
1025	if (empty) {
1026		spin_unlock(&ino->i_lock);
1027		dprintk("NFS: %s no layout segments to return\n", __func__);
1028		goto out_put_layout_hdr;
1029	}
1030
1031	set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1032	send = pnfs_prepare_layoutreturn(lo);
1033	spin_unlock(&ino->i_lock);
1034	pnfs_free_lseg_list(&tmp_list);
1035	if (send)
1036		status = pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1037out_put_layout_hdr:
1038	pnfs_put_layout_hdr(lo);
1039out:
1040	dprintk("<-- %s status: %d\n", __func__, status);
1041	return status;
1042}
1043EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1044
1045int
1046pnfs_commit_and_return_layout(struct inode *inode)
1047{
1048	struct pnfs_layout_hdr *lo;
1049	int ret;
1050
1051	spin_lock(&inode->i_lock);
1052	lo = NFS_I(inode)->layout;
1053	if (lo == NULL) {
1054		spin_unlock(&inode->i_lock);
1055		return 0;
1056	}
1057	pnfs_get_layout_hdr(lo);
1058	/* Block new layoutgets and read/write to ds */
1059	lo->plh_block_lgets++;
1060	spin_unlock(&inode->i_lock);
1061	filemap_fdatawait(inode->i_mapping);
1062	ret = pnfs_layoutcommit_inode(inode, true);
1063	if (ret == 0)
1064		ret = _pnfs_return_layout(inode);
1065	spin_lock(&inode->i_lock);
1066	lo->plh_block_lgets--;
1067	spin_unlock(&inode->i_lock);
1068	pnfs_put_layout_hdr(lo);
1069	return ret;
1070}
1071
1072bool pnfs_roc(struct inode *ino)
1073{
1074	struct nfs_inode *nfsi = NFS_I(ino);
1075	struct nfs_open_context *ctx;
1076	struct nfs4_state *state;
1077	struct pnfs_layout_hdr *lo;
1078	struct pnfs_layout_segment *lseg, *tmp;
1079	nfs4_stateid stateid;
1080	LIST_HEAD(tmp_list);
1081	bool found = false, layoutreturn = false, roc = false;
1082
1083	spin_lock(&ino->i_lock);
1084	lo = nfsi->layout;
1085	if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1086		goto out_noroc;
1087
1088	/* no roc if we hold a delegation */
1089	if (nfs4_check_delegation(ino, FMODE_READ))
1090		goto out_noroc;
1091
1092	list_for_each_entry(ctx, &nfsi->open_files, list) {
1093		state = ctx->state;
1094		/* Don't return layout if there is open file state */
1095		if (state != NULL && state->state != 0)
1096			goto out_noroc;
1097	}
1098
1099	stateid = lo->plh_stateid;
1100	/* always send layoutreturn if being marked so */
1101	if (test_and_clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1102				   &lo->plh_flags))
1103		layoutreturn = pnfs_prepare_layoutreturn(lo);
1104
1105	pnfs_clear_retry_layoutget(lo);
1106	list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1107		/* If we are sending layoutreturn, invalidate all valid lsegs */
1108		if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1109			mark_lseg_invalid(lseg, &tmp_list);
1110			found = true;
1111		}
1112	/* ROC in two conditions:
1113	 * 1. there are ROC lsegs
1114	 * 2. we don't send layoutreturn
1115	 */
1116	if (found && !layoutreturn) {
1117		/* lo ref dropped in pnfs_roc_release() */
1118		pnfs_get_layout_hdr(lo);
1119		roc = true;
1120	}
1121
1122out_noroc:
1123	spin_unlock(&ino->i_lock);
1124	pnfs_free_lseg_list(&tmp_list);
1125	pnfs_layoutcommit_inode(ino, true);
1126	if (layoutreturn)
1127		pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1128	return roc;
1129}
1130
1131void pnfs_roc_release(struct inode *ino)
1132{
1133	struct pnfs_layout_hdr *lo;
1134
1135	spin_lock(&ino->i_lock);
1136	lo = NFS_I(ino)->layout;
1137	pnfs_clear_layoutreturn_waitbit(lo);
1138	if (atomic_dec_and_test(&lo->plh_refcount)) {
1139		pnfs_detach_layout_hdr(lo);
1140		spin_unlock(&ino->i_lock);
1141		pnfs_free_layout_hdr(lo);
1142	} else
1143		spin_unlock(&ino->i_lock);
1144}
1145
1146void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1147{
1148	struct pnfs_layout_hdr *lo;
1149
1150	spin_lock(&ino->i_lock);
1151	lo = NFS_I(ino)->layout;
1152	if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1153		lo->plh_barrier = barrier;
1154	spin_unlock(&ino->i_lock);
1155	trace_nfs4_layoutreturn_on_close(ino, 0);
1156}
1157
1158void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
1159{
1160	struct nfs_inode *nfsi = NFS_I(ino);
1161	struct pnfs_layout_hdr *lo;
1162	u32 current_seqid;
1163
1164	spin_lock(&ino->i_lock);
1165	lo = nfsi->layout;
1166	current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1167
1168	/* Since close does not return a layout stateid for use as
1169	 * a barrier, we choose the worst-case barrier.
1170	 */
1171	*barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1172	spin_unlock(&ino->i_lock);
1173}
1174
1175bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1176{
1177	struct nfs_inode *nfsi = NFS_I(ino);
1178        struct pnfs_layout_hdr *lo;
1179        bool sleep = false;
1180
1181	/* we might not have grabbed lo reference. so need to check under
1182	 * i_lock */
1183        spin_lock(&ino->i_lock);
1184        lo = nfsi->layout;
1185        if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
1186                sleep = true;
1187        spin_unlock(&ino->i_lock);
1188
1189        if (sleep)
1190                rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1191
1192        return sleep;
1193}
1194
1195/*
1196 * Compare two layout segments for sorting into layout cache.
1197 * We want to preferentially return RW over RO layouts, so ensure those
1198 * are seen first.
1199 */
1200static s64
1201pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1202	   const struct pnfs_layout_range *l2)
1203{
1204	s64 d;
1205
1206	/* high offset > low offset */
1207	d = l1->offset - l2->offset;
1208	if (d)
1209		return d;
1210
1211	/* short length > long length */
1212	d = l2->length - l1->length;
1213	if (d)
1214		return d;
1215
1216	/* read > read/write */
1217	return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1218}
1219
1220static bool
1221pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1222		const struct pnfs_layout_range *l2)
1223{
1224	return pnfs_lseg_range_cmp(l1, l2) > 0;
1225}
1226
1227static bool
1228pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1229		struct pnfs_layout_segment *old)
1230{
1231	return false;
1232}
1233
1234void
1235pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1236		   struct pnfs_layout_segment *lseg,
1237		   bool (*is_after)(const struct pnfs_layout_range *,
1238			   const struct pnfs_layout_range *),
1239		   bool (*do_merge)(struct pnfs_layout_segment *,
1240			   struct pnfs_layout_segment *),
1241		   struct list_head *free_me)
1242{
1243	struct pnfs_layout_segment *lp, *tmp;
1244
1245	dprintk("%s:Begin\n", __func__);
1246
1247	list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1248		if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1249			continue;
1250		if (do_merge(lseg, lp)) {
1251			mark_lseg_invalid(lp, free_me);
1252			continue;
1253		}
1254		if (is_after(&lseg->pls_range, &lp->pls_range))
1255			continue;
1256		list_add_tail(&lseg->pls_list, &lp->pls_list);
1257		dprintk("%s: inserted lseg %p "
1258			"iomode %d offset %llu length %llu before "
1259			"lp %p iomode %d offset %llu length %llu\n",
1260			__func__, lseg, lseg->pls_range.iomode,
1261			lseg->pls_range.offset, lseg->pls_range.length,
1262			lp, lp->pls_range.iomode, lp->pls_range.offset,
1263			lp->pls_range.length);
1264		goto out;
1265	}
1266	list_add_tail(&lseg->pls_list, &lo->plh_segs);
1267	dprintk("%s: inserted lseg %p "
1268		"iomode %d offset %llu length %llu at tail\n",
1269		__func__, lseg, lseg->pls_range.iomode,
1270		lseg->pls_range.offset, lseg->pls_range.length);
1271out:
1272	pnfs_get_layout_hdr(lo);
1273
1274	dprintk("%s:Return\n", __func__);
1275}
1276EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1277
1278static void
1279pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1280		   struct pnfs_layout_segment *lseg,
1281		   struct list_head *free_me)
1282{
1283	struct inode *inode = lo->plh_inode;
1284	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1285
1286	if (ld->add_lseg != NULL)
1287		ld->add_lseg(lo, lseg, free_me);
1288	else
1289		pnfs_generic_layout_insert_lseg(lo, lseg,
1290				pnfs_lseg_range_is_after,
1291				pnfs_lseg_no_merge,
1292				free_me);
1293}
1294
1295static struct pnfs_layout_hdr *
1296alloc_init_layout_hdr(struct inode *ino,
1297		      struct nfs_open_context *ctx,
1298		      gfp_t gfp_flags)
1299{
1300	struct pnfs_layout_hdr *lo;
1301
1302	lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1303	if (!lo)
1304		return NULL;
1305	atomic_set(&lo->plh_refcount, 1);
1306	INIT_LIST_HEAD(&lo->plh_layouts);
1307	INIT_LIST_HEAD(&lo->plh_segs);
1308	INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1309	lo->plh_inode = ino;
1310	lo->plh_lc_cred = get_rpccred(ctx->cred);
1311	return lo;
1312}
1313
1314static struct pnfs_layout_hdr *
1315pnfs_find_alloc_layout(struct inode *ino,
1316		       struct nfs_open_context *ctx,
1317		       gfp_t gfp_flags)
1318{
1319	struct nfs_inode *nfsi = NFS_I(ino);
1320	struct pnfs_layout_hdr *new = NULL;
1321
1322	dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1323
1324	if (nfsi->layout != NULL)
1325		goto out_existing;
1326	spin_unlock(&ino->i_lock);
1327	new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1328	spin_lock(&ino->i_lock);
1329
1330	if (likely(nfsi->layout == NULL)) {	/* Won the race? */
1331		nfsi->layout = new;
1332		return new;
1333	} else if (new != NULL)
1334		pnfs_free_layout_hdr(new);
1335out_existing:
1336	pnfs_get_layout_hdr(nfsi->layout);
1337	return nfsi->layout;
1338}
1339
1340/*
1341 * iomode matching rules:
1342 * iomode	lseg	match
1343 * -----	-----	-----
1344 * ANY		READ	true
1345 * ANY		RW	true
1346 * RW		READ	false
1347 * RW		RW	true
1348 * READ		READ	true
1349 * READ		RW	true
1350 */
1351static bool
1352pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1353		 const struct pnfs_layout_range *range)
1354{
1355	struct pnfs_layout_range range1;
1356
1357	if ((range->iomode == IOMODE_RW &&
1358	     ls_range->iomode != IOMODE_RW) ||
1359	    !pnfs_lseg_range_intersecting(ls_range, range))
1360		return 0;
1361
1362	/* range1 covers only the first byte in the range */
1363	range1 = *range;
1364	range1.length = 1;
1365	return pnfs_lseg_range_contained(ls_range, &range1);
1366}
1367
1368/*
1369 * lookup range in layout
1370 */
1371static struct pnfs_layout_segment *
1372pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1373		struct pnfs_layout_range *range)
1374{
1375	struct pnfs_layout_segment *lseg, *ret = NULL;
1376
1377	dprintk("%s:Begin\n", __func__);
1378
1379	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1380		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1381		    !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1382		    pnfs_lseg_range_match(&lseg->pls_range, range)) {
1383			ret = pnfs_get_lseg(lseg);
1384			break;
1385		}
1386	}
1387
1388	dprintk("%s:Return lseg %p ref %d\n",
1389		__func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1390	return ret;
1391}
1392
1393/*
1394 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1395 * to the MDS or over pNFS
1396 *
1397 * The nfs_inode read_io and write_io fields are cumulative counters reset
1398 * when there are no layout segments. Note that in pnfs_update_layout iomode
1399 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1400 * WRITE request.
1401 *
1402 * A return of true means use MDS I/O.
1403 *
1404 * From rfc 5661:
1405 * If a file's size is smaller than the file size threshold, data accesses
1406 * SHOULD be sent to the metadata server.  If an I/O request has a length that
1407 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1408 * server.  If both file size and I/O size are provided, the client SHOULD
1409 * reach or exceed  both thresholds before sending its read or write
1410 * requests to the data server.
1411 */
1412static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1413				     struct inode *ino, int iomode)
1414{
1415	struct nfs4_threshold *t = ctx->mdsthreshold;
1416	struct nfs_inode *nfsi = NFS_I(ino);
1417	loff_t fsize = i_size_read(ino);
1418	bool size = false, size_set = false, io = false, io_set = false, ret = false;
1419
1420	if (t == NULL)
1421		return ret;
1422
1423	dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1424		__func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1425
1426	switch (iomode) {
1427	case IOMODE_READ:
1428		if (t->bm & THRESHOLD_RD) {
1429			dprintk("%s fsize %llu\n", __func__, fsize);
1430			size_set = true;
1431			if (fsize < t->rd_sz)
1432				size = true;
1433		}
1434		if (t->bm & THRESHOLD_RD_IO) {
1435			dprintk("%s nfsi->read_io %llu\n", __func__,
1436				nfsi->read_io);
1437			io_set = true;
1438			if (nfsi->read_io < t->rd_io_sz)
1439				io = true;
1440		}
1441		break;
1442	case IOMODE_RW:
1443		if (t->bm & THRESHOLD_WR) {
1444			dprintk("%s fsize %llu\n", __func__, fsize);
1445			size_set = true;
1446			if (fsize < t->wr_sz)
1447				size = true;
1448		}
1449		if (t->bm & THRESHOLD_WR_IO) {
1450			dprintk("%s nfsi->write_io %llu\n", __func__,
1451				nfsi->write_io);
1452			io_set = true;
1453			if (nfsi->write_io < t->wr_io_sz)
1454				io = true;
1455		}
1456		break;
1457	}
1458	if (size_set && io_set) {
1459		if (size && io)
1460			ret = true;
1461	} else if (size || io)
1462		ret = true;
1463
1464	dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1465	return ret;
1466}
1467
1468/* stop waiting if someone clears NFS_LAYOUT_RETRY_LAYOUTGET bit. */
1469static int pnfs_layoutget_retry_bit_wait(struct wait_bit_key *key, int mode)
1470{
1471	if (!test_bit(NFS_LAYOUT_RETRY_LAYOUTGET, key->flags))
1472		return 1;
1473	return nfs_wait_bit_killable(key, mode);
1474}
1475
1476static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1477{
1478	if (!pnfs_should_retry_layoutget(lo))
1479		return false;
1480	/*
1481	 * send layoutcommit as it can hold up layoutreturn due to lseg
1482	 * reference
1483	 */
1484	pnfs_layoutcommit_inode(lo->plh_inode, false);
1485	return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1486				   pnfs_layoutget_retry_bit_wait,
1487				   TASK_UNINTERRUPTIBLE);
1488}
1489
1490static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1491{
1492	unsigned long *bitlock = &lo->plh_flags;
1493
1494	clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1495	smp_mb__after_atomic();
1496	wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1497}
1498
1499/*
1500 * Layout segment is retreived from the server if not cached.
1501 * The appropriate layout segment is referenced and returned to the caller.
1502 */
1503struct pnfs_layout_segment *
1504pnfs_update_layout(struct inode *ino,
1505		   struct nfs_open_context *ctx,
1506		   loff_t pos,
1507		   u64 count,
1508		   enum pnfs_iomode iomode,
1509		   gfp_t gfp_flags)
1510{
1511	struct pnfs_layout_range arg = {
1512		.iomode = iomode,
1513		.offset = pos,
1514		.length = count,
1515	};
1516	unsigned pg_offset;
1517	struct nfs_server *server = NFS_SERVER(ino);
1518	struct nfs_client *clp = server->nfs_client;
1519	struct pnfs_layout_hdr *lo;
1520	struct pnfs_layout_segment *lseg = NULL;
1521	bool first;
1522
1523	if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1524		goto out;
1525
1526	if (iomode == IOMODE_READ && i_size_read(ino) == 0)
1527		goto out;
1528
1529	if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1530		goto out;
1531
1532lookup_again:
1533	first = false;
1534	spin_lock(&ino->i_lock);
1535	lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1536	if (lo == NULL) {
1537		spin_unlock(&ino->i_lock);
1538		goto out;
1539	}
1540
1541	/* Do we even need to bother with this? */
1542	if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1543		dprintk("%s matches recall, use MDS\n", __func__);
1544		goto out_unlock;
1545	}
1546
1547	/* if LAYOUTGET already failed once we don't try again */
1548	if (pnfs_layout_io_test_failed(lo, iomode) &&
1549	    !pnfs_should_retry_layoutget(lo))
1550		goto out_unlock;
1551
1552	first = list_empty(&lo->plh_segs);
1553	if (first) {
1554		/* The first layoutget for the file. Need to serialize per
1555		 * RFC 5661 Errata 3208.
1556		 */
1557		if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1558				     &lo->plh_flags)) {
1559			spin_unlock(&ino->i_lock);
1560			wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1561				    TASK_UNINTERRUPTIBLE);
1562			pnfs_put_layout_hdr(lo);
1563			goto lookup_again;
1564		}
1565	} else {
1566		/* Check to see if the layout for the given range
1567		 * already exists
1568		 */
1569		lseg = pnfs_find_lseg(lo, &arg);
1570		if (lseg)
1571			goto out_unlock;
1572	}
1573
1574	/*
1575	 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1576	 * for LAYOUTRETURN even if first is true.
1577	 */
1578	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1579		spin_unlock(&ino->i_lock);
1580		dprintk("%s wait for layoutreturn\n", __func__);
1581		if (pnfs_prepare_to_retry_layoutget(lo)) {
1582			if (first)
1583				pnfs_clear_first_layoutget(lo);
1584			pnfs_put_layout_hdr(lo);
1585			dprintk("%s retrying\n", __func__);
1586			goto lookup_again;
1587		}
1588		goto out_put_layout_hdr;
1589	}
1590
1591	if (pnfs_layoutgets_blocked(lo))
1592		goto out_unlock;
1593	atomic_inc(&lo->plh_outstanding);
1594	spin_unlock(&ino->i_lock);
1595
1596	if (list_empty(&lo->plh_layouts)) {
1597		/* The lo must be on the clp list if there is any
1598		 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1599		 */
1600		spin_lock(&clp->cl_lock);
1601		if (list_empty(&lo->plh_layouts))
1602			list_add_tail(&lo->plh_layouts, &server->layouts);
1603		spin_unlock(&clp->cl_lock);
1604	}
1605
1606	pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1607	if (pg_offset) {
1608		arg.offset -= pg_offset;
1609		arg.length += pg_offset;
1610	}
1611	if (arg.length != NFS4_MAX_UINT64)
1612		arg.length = PAGE_CACHE_ALIGN(arg.length);
1613
1614	lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1615	pnfs_clear_retry_layoutget(lo);
1616	atomic_dec(&lo->plh_outstanding);
1617out_put_layout_hdr:
1618	if (first)
1619		pnfs_clear_first_layoutget(lo);
1620	pnfs_put_layout_hdr(lo);
1621out:
1622	dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1623			"(%s, offset: %llu, length: %llu)\n",
1624			__func__, ino->i_sb->s_id,
1625			(unsigned long long)NFS_FILEID(ino),
1626			lseg == NULL ? "not found" : "found",
1627			iomode==IOMODE_RW ?  "read/write" : "read-only",
1628			(unsigned long long)pos,
1629			(unsigned long long)count);
1630	return lseg;
1631out_unlock:
1632	spin_unlock(&ino->i_lock);
1633	goto out_put_layout_hdr;
1634}
1635EXPORT_SYMBOL_GPL(pnfs_update_layout);
1636
1637static bool
1638pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1639{
1640	switch (range->iomode) {
1641	case IOMODE_READ:
1642	case IOMODE_RW:
1643		break;
1644	default:
1645		return false;
1646	}
1647	if (range->offset == NFS4_MAX_UINT64)
1648		return false;
1649	if (range->length == 0)
1650		return false;
1651	if (range->length != NFS4_MAX_UINT64 &&
1652	    range->length > NFS4_MAX_UINT64 - range->offset)
1653		return false;
1654	return true;
1655}
1656
1657struct pnfs_layout_segment *
1658pnfs_layout_process(struct nfs4_layoutget *lgp)
1659{
1660	struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1661	struct nfs4_layoutget_res *res = &lgp->res;
1662	struct pnfs_layout_segment *lseg;
1663	struct inode *ino = lo->plh_inode;
1664	LIST_HEAD(free_me);
1665	int status = -EINVAL;
1666
1667	if (!pnfs_sanity_check_layout_range(&res->range))
1668		goto out;
1669
1670	/* Inject layout blob into I/O device driver */
1671	lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1672	if (!lseg || IS_ERR(lseg)) {
1673		if (!lseg)
1674			status = -ENOMEM;
1675		else
1676			status = PTR_ERR(lseg);
1677		dprintk("%s: Could not allocate layout: error %d\n",
1678		       __func__, status);
1679		goto out;
1680	}
1681
1682	init_lseg(lo, lseg);
1683	lseg->pls_range = res->range;
1684
1685	spin_lock(&ino->i_lock);
1686	if (pnfs_layoutgets_blocked(lo)) {
1687		dprintk("%s forget reply due to state\n", __func__);
1688		goto out_forget_reply;
1689	}
1690
1691	if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1692		/* existing state ID, make sure the sequence number matches. */
1693		if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1694			dprintk("%s forget reply due to sequence\n", __func__);
1695			status = -EAGAIN;
1696			goto out_forget_reply;
1697		}
1698		pnfs_set_layout_stateid(lo, &res->stateid, false);
1699	} else {
1700		/*
1701		 * We got an entirely new state ID.  Mark all segments for the
1702		 * inode invalid, and don't bother validating the stateid
1703		 * sequence number.
1704		 */
1705		pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1706
1707		nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1708		lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1709	}
1710
1711	clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1712
1713	pnfs_get_lseg(lseg);
1714	pnfs_layout_insert_lseg(lo, lseg, &free_me);
1715
1716	if (res->return_on_close)
1717		set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1718
1719	spin_unlock(&ino->i_lock);
1720	pnfs_free_lseg_list(&free_me);
1721	return lseg;
1722out:
1723	return ERR_PTR(status);
1724
1725out_forget_reply:
1726	spin_unlock(&ino->i_lock);
1727	lseg->pls_layout = lo;
1728	NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1729	goto out;
1730}
1731
1732static void
1733pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1734				struct list_head *tmp_list,
1735				struct pnfs_layout_range *return_range)
1736{
1737	struct pnfs_layout_segment *lseg, *next;
1738
1739	dprintk("%s:Begin lo %p\n", __func__, lo);
1740
1741	if (list_empty(&lo->plh_segs))
1742		return;
1743
1744	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1745		if (should_free_lseg(&lseg->pls_range, return_range)) {
1746			dprintk("%s: marking lseg %p iomode %d "
1747				"offset %llu length %llu\n", __func__,
1748				lseg, lseg->pls_range.iomode,
1749				lseg->pls_range.offset,
1750				lseg->pls_range.length);
1751			set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1752			mark_lseg_invalid(lseg, tmp_list);
1753			set_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1754					&lo->plh_flags);
1755		}
1756}
1757
1758void pnfs_error_mark_layout_for_return(struct inode *inode,
1759				       struct pnfs_layout_segment *lseg)
1760{
1761	struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1762	int iomode = pnfs_iomode_to_fail_bit(lseg->pls_range.iomode);
1763	struct pnfs_layout_range range = {
1764		.iomode = lseg->pls_range.iomode,
1765		.offset = 0,
1766		.length = NFS4_MAX_UINT64,
1767	};
1768	LIST_HEAD(free_me);
1769
1770	spin_lock(&inode->i_lock);
1771	/* set failure bit so that pnfs path will be retried later */
1772	pnfs_layout_set_fail_bit(lo, iomode);
1773	if (lo->plh_return_iomode == 0)
1774		lo->plh_return_iomode = range.iomode;
1775	else if (lo->plh_return_iomode != range.iomode)
1776		lo->plh_return_iomode = IOMODE_ANY;
1777	/*
1778	 * mark all matching lsegs so that we are sure to have no live
1779	 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1780	 * for how it works.
1781	 */
1782	pnfs_mark_matching_lsegs_return(lo, &free_me, &range);
1783	spin_unlock(&inode->i_lock);
1784	pnfs_free_lseg_list(&free_me);
1785}
1786EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1787
1788void
1789pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1790{
1791	u64 rd_size = req->wb_bytes;
1792
1793	if (pgio->pg_lseg == NULL) {
1794		if (pgio->pg_dreq == NULL)
1795			rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1796		else
1797			rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1798
1799		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1800						   req->wb_context,
1801						   req_offset(req),
1802						   rd_size,
1803						   IOMODE_READ,
1804						   GFP_KERNEL);
1805	}
1806	/* If no lseg, fall back to read through mds */
1807	if (pgio->pg_lseg == NULL)
1808		nfs_pageio_reset_read_mds(pgio);
1809
1810}
1811EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1812
1813void
1814pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1815			   struct nfs_page *req, u64 wb_size)
1816{
1817	if (pgio->pg_lseg == NULL)
1818		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1819						   req->wb_context,
1820						   req_offset(req),
1821						   wb_size,
1822						   IOMODE_RW,
1823						   GFP_NOFS);
1824	/* If no lseg, fall back to write through mds */
1825	if (pgio->pg_lseg == NULL)
1826		nfs_pageio_reset_write_mds(pgio);
1827}
1828EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1829
1830void
1831pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1832{
1833	if (desc->pg_lseg) {
1834		pnfs_put_lseg(desc->pg_lseg);
1835		desc->pg_lseg = NULL;
1836	}
1837}
1838EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1839
1840/*
1841 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1842 * of bytes (maximum @req->wb_bytes) that can be coalesced.
1843 */
1844size_t
1845pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1846		     struct nfs_page *prev, struct nfs_page *req)
1847{
1848	unsigned int size;
1849	u64 seg_end, req_start, seg_left;
1850
1851	size = nfs_generic_pg_test(pgio, prev, req);
1852	if (!size)
1853		return 0;
1854
1855	/*
1856	 * 'size' contains the number of bytes left in the current page (up
1857	 * to the original size asked for in @req->wb_bytes).
1858	 *
1859	 * Calculate how many bytes are left in the layout segment
1860	 * and if there are less bytes than 'size', return that instead.
1861	 *
1862	 * Please also note that 'end_offset' is actually the offset of the
1863	 * first byte that lies outside the pnfs_layout_range. FIXME?
1864	 *
1865	 */
1866	if (pgio->pg_lseg) {
1867		seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1868				     pgio->pg_lseg->pls_range.length);
1869		req_start = req_offset(req);
1870		WARN_ON_ONCE(req_start >= seg_end);
1871		/* start of request is past the last byte of this segment */
1872		if (req_start >= seg_end) {
1873			/* reference the new lseg */
1874			if (pgio->pg_ops->pg_cleanup)
1875				pgio->pg_ops->pg_cleanup(pgio);
1876			if (pgio->pg_ops->pg_init)
1877				pgio->pg_ops->pg_init(pgio, req);
1878			return 0;
1879		}
1880
1881		/* adjust 'size' iff there are fewer bytes left in the
1882		 * segment than what nfs_generic_pg_test returned */
1883		seg_left = seg_end - req_start;
1884		if (seg_left < size)
1885			size = (unsigned int)seg_left;
1886	}
1887
1888	return size;
1889}
1890EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1891
1892int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1893{
1894	struct nfs_pageio_descriptor pgio;
1895
1896	/* Resend all requests through the MDS */
1897	nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1898			      hdr->completion_ops);
1899	set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
1900	return nfs_pageio_resend(&pgio, hdr);
1901}
1902EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1903
1904static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1905{
1906
1907	dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1908	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1909	    PNFS_LAYOUTRET_ON_ERROR) {
1910		pnfs_return_layout(hdr->inode);
1911	}
1912	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1913		hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1914}
1915
1916/*
1917 * Called by non rpc-based layout drivers
1918 */
1919void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1920{
1921	if (likely(!hdr->pnfs_error)) {
1922		pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1923				hdr->mds_offset + hdr->res.count);
1924		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1925	}
1926	trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1927	if (unlikely(hdr->pnfs_error))
1928		pnfs_ld_handle_write_error(hdr);
1929	hdr->mds_ops->rpc_release(hdr);
1930}
1931EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1932
1933static void
1934pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1935		struct nfs_pgio_header *hdr)
1936{
1937	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1938
1939	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1940		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1941		nfs_pageio_reset_write_mds(desc);
1942		mirror->pg_recoalesce = 1;
1943	}
1944	nfs_pgio_data_destroy(hdr);
1945	hdr->release(hdr);
1946}
1947
1948static enum pnfs_try_status
1949pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1950			const struct rpc_call_ops *call_ops,
1951			struct pnfs_layout_segment *lseg,
1952			int how)
1953{
1954	struct inode *inode = hdr->inode;
1955	enum pnfs_try_status trypnfs;
1956	struct nfs_server *nfss = NFS_SERVER(inode);
1957
1958	hdr->mds_ops = call_ops;
1959
1960	dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1961		inode->i_ino, hdr->args.count, hdr->args.offset, how);
1962	trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1963	if (trypnfs != PNFS_NOT_ATTEMPTED)
1964		nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1965	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1966	return trypnfs;
1967}
1968
1969static void
1970pnfs_do_write(struct nfs_pageio_descriptor *desc,
1971	      struct nfs_pgio_header *hdr, int how)
1972{
1973	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1974	struct pnfs_layout_segment *lseg = desc->pg_lseg;
1975	enum pnfs_try_status trypnfs;
1976
1977	trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1978	if (trypnfs == PNFS_NOT_ATTEMPTED)
1979		pnfs_write_through_mds(desc, hdr);
1980}
1981
1982static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1983{
1984	pnfs_put_lseg(hdr->lseg);
1985	nfs_pgio_header_free(hdr);
1986}
1987
1988int
1989pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1990{
1991	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1992
1993	struct nfs_pgio_header *hdr;
1994	int ret;
1995
1996	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1997	if (!hdr) {
1998		desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
1999		return -ENOMEM;
2000	}
2001	nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2002
2003	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2004	ret = nfs_generic_pgio(desc, hdr);
2005	if (!ret)
2006		pnfs_do_write(desc, hdr, desc->pg_ioflags);
2007
2008	return ret;
2009}
2010EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2011
2012int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2013{
2014	struct nfs_pageio_descriptor pgio;
2015
2016	/* Resend all requests through the MDS */
2017	nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2018	return nfs_pageio_resend(&pgio, hdr);
2019}
2020EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2021
2022static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2023{
2024	dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2025	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2026	    PNFS_LAYOUTRET_ON_ERROR) {
2027		pnfs_return_layout(hdr->inode);
2028	}
2029	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2030		hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2031}
2032
2033/*
2034 * Called by non rpc-based layout drivers
2035 */
2036void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2037{
2038	if (likely(!hdr->pnfs_error)) {
2039		__nfs4_read_done_cb(hdr);
2040		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2041	}
2042	trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2043	if (unlikely(hdr->pnfs_error))
2044		pnfs_ld_handle_read_error(hdr);
2045	hdr->mds_ops->rpc_release(hdr);
2046}
2047EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2048
2049static void
2050pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2051		struct nfs_pgio_header *hdr)
2052{
2053	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2054
2055	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2056		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2057		nfs_pageio_reset_read_mds(desc);
2058		mirror->pg_recoalesce = 1;
2059	}
2060	nfs_pgio_data_destroy(hdr);
2061	hdr->release(hdr);
2062}
2063
2064/*
2065 * Call the appropriate parallel I/O subsystem read function.
2066 */
2067static enum pnfs_try_status
2068pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2069		       const struct rpc_call_ops *call_ops,
2070		       struct pnfs_layout_segment *lseg)
2071{
2072	struct inode *inode = hdr->inode;
2073	struct nfs_server *nfss = NFS_SERVER(inode);
2074	enum pnfs_try_status trypnfs;
2075
2076	hdr->mds_ops = call_ops;
2077
2078	dprintk("%s: Reading ino:%lu %u@%llu\n",
2079		__func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2080
2081	trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2082	if (trypnfs != PNFS_NOT_ATTEMPTED)
2083		nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2084	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2085	return trypnfs;
2086}
2087
2088/* Resend all requests through pnfs. */
2089int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2090{
2091	struct nfs_pageio_descriptor pgio;
2092
2093	nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
2094	return nfs_pageio_resend(&pgio, hdr);
2095}
2096EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2097
2098static void
2099pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2100{
2101	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2102	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2103	enum pnfs_try_status trypnfs;
2104	int err = 0;
2105
2106	trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2107	if (trypnfs == PNFS_TRY_AGAIN)
2108		err = pnfs_read_resend_pnfs(hdr);
2109	if (trypnfs == PNFS_NOT_ATTEMPTED || err)
2110		pnfs_read_through_mds(desc, hdr);
2111}
2112
2113static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2114{
2115	pnfs_put_lseg(hdr->lseg);
2116	nfs_pgio_header_free(hdr);
2117}
2118
2119int
2120pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2121{
2122	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2123
2124	struct nfs_pgio_header *hdr;
2125	int ret;
2126
2127	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2128	if (!hdr) {
2129		desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
2130		return -ENOMEM;
2131	}
2132	nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2133	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2134	ret = nfs_generic_pgio(desc, hdr);
2135	if (!ret)
2136		pnfs_do_read(desc, hdr);
2137	return ret;
2138}
2139EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2140
2141static void pnfs_clear_layoutcommitting(struct inode *inode)
2142{
2143	unsigned long *bitlock = &NFS_I(inode)->flags;
2144
2145	clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2146	smp_mb__after_atomic();
2147	wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2148}
2149
2150/*
2151 * There can be multiple RW segments.
2152 */
2153static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2154{
2155	struct pnfs_layout_segment *lseg;
2156
2157	list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2158		if (lseg->pls_range.iomode == IOMODE_RW &&
2159		    test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2160			list_add(&lseg->pls_lc_list, listp);
2161	}
2162}
2163
2164static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2165{
2166	struct pnfs_layout_segment *lseg, *tmp;
2167
2168	/* Matched by references in pnfs_set_layoutcommit */
2169	list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2170		list_del_init(&lseg->pls_lc_list);
2171		pnfs_put_lseg(lseg);
2172	}
2173
2174	pnfs_clear_layoutcommitting(inode);
2175}
2176
2177void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2178{
2179	pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2180}
2181EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2182
2183void
2184pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2185		loff_t end_pos)
2186{
2187	struct nfs_inode *nfsi = NFS_I(inode);
2188	bool mark_as_dirty = false;
2189
2190	spin_lock(&inode->i_lock);
2191	if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2192		nfsi->layout->plh_lwb = end_pos;
2193		mark_as_dirty = true;
2194		dprintk("%s: Set layoutcommit for inode %lu ",
2195			__func__, inode->i_ino);
2196	} else if (end_pos > nfsi->layout->plh_lwb)
2197		nfsi->layout->plh_lwb = end_pos;
2198	if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2199		/* references matched in nfs4_layoutcommit_release */
2200		pnfs_get_lseg(lseg);
2201	}
2202	spin_unlock(&inode->i_lock);
2203	dprintk("%s: lseg %p end_pos %llu\n",
2204		__func__, lseg, nfsi->layout->plh_lwb);
2205
2206	/* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2207	 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2208	if (mark_as_dirty)
2209		mark_inode_dirty_sync(inode);
2210}
2211EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2212
2213void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2214{
2215	struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2216
2217	if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2218		nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2219	pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2220}
2221
2222/*
2223 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2224 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2225 * data to disk to allow the server to recover the data if it crashes.
2226 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2227 * is off, and a COMMIT is sent to a data server, or
2228 * if WRITEs to a data server return NFS_DATA_SYNC.
2229 */
2230int
2231pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2232{
2233	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2234	struct nfs4_layoutcommit_data *data;
2235	struct nfs_inode *nfsi = NFS_I(inode);
2236	loff_t end_pos;
2237	int status;
2238
2239	if (!pnfs_layoutcommit_outstanding(inode))
2240		return 0;
2241
2242	dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2243
2244	status = -EAGAIN;
2245	if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2246		if (!sync)
2247			goto out;
2248		status = wait_on_bit_lock_action(&nfsi->flags,
2249				NFS_INO_LAYOUTCOMMITTING,
2250				nfs_wait_bit_killable,
2251				TASK_KILLABLE);
2252		if (status)
2253			goto out;
2254	}
2255
2256	status = -ENOMEM;
2257	/* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2258	data = kzalloc(sizeof(*data), GFP_NOFS);
2259	if (!data)
2260		goto clear_layoutcommitting;
2261
2262	status = 0;
2263	spin_lock(&inode->i_lock);
2264	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2265		goto out_unlock;
2266
2267	INIT_LIST_HEAD(&data->lseg_list);
2268	pnfs_list_write_lseg(inode, &data->lseg_list);
2269
2270	end_pos = nfsi->layout->plh_lwb;
2271
2272	nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2273	spin_unlock(&inode->i_lock);
2274
2275	data->args.inode = inode;
2276	data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2277	nfs_fattr_init(&data->fattr);
2278	data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2279	data->res.fattr = &data->fattr;
2280	data->args.lastbytewritten = end_pos - 1;
2281	data->res.server = NFS_SERVER(inode);
2282
2283	if (ld->prepare_layoutcommit) {
2284		status = ld->prepare_layoutcommit(&data->args);
2285		if (status) {
2286			put_rpccred(data->cred);
2287			spin_lock(&inode->i_lock);
2288			set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2289			if (end_pos > nfsi->layout->plh_lwb)
2290				nfsi->layout->plh_lwb = end_pos;
2291			goto out_unlock;
2292		}
2293	}
2294
2295
2296	status = nfs4_proc_layoutcommit(data, sync);
2297out:
2298	if (status)
2299		mark_inode_dirty_sync(inode);
2300	dprintk("<-- %s status %d\n", __func__, status);
2301	return status;
2302out_unlock:
2303	spin_unlock(&inode->i_lock);
2304	kfree(data);
2305clear_layoutcommitting:
2306	pnfs_clear_layoutcommitting(inode);
2307	goto out;
2308}
2309EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2310
2311int
2312pnfs_generic_sync(struct inode *inode, bool datasync)
2313{
2314	return pnfs_layoutcommit_inode(inode, true);
2315}
2316EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2317
2318struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2319{
2320	struct nfs4_threshold *thp;
2321
2322	thp = kzalloc(sizeof(*thp), GFP_NOFS);
2323	if (!thp) {
2324		dprintk("%s mdsthreshold allocation failed\n", __func__);
2325		return NULL;
2326	}
2327	return thp;
2328}
2329
2330#if IS_ENABLED(CONFIG_NFS_V4_2)
2331int
2332pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2333{
2334	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2335	struct nfs_server *server = NFS_SERVER(inode);
2336	struct nfs_inode *nfsi = NFS_I(inode);
2337	struct nfs42_layoutstat_data *data;
2338	struct pnfs_layout_hdr *hdr;
2339	int status = 0;
2340
2341	if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2342		goto out;
2343
2344	if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2345		goto out;
2346
2347	if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2348		goto out;
2349
2350	spin_lock(&inode->i_lock);
2351	if (!NFS_I(inode)->layout) {
2352		spin_unlock(&inode->i_lock);
2353		goto out;
2354	}
2355	hdr = NFS_I(inode)->layout;
2356	pnfs_get_layout_hdr(hdr);
2357	spin_unlock(&inode->i_lock);
2358
2359	data = kzalloc(sizeof(*data), gfp_flags);
2360	if (!data) {
2361		status = -ENOMEM;
2362		goto out_put;
2363	}
2364
2365	data->args.fh = NFS_FH(inode);
2366	data->args.inode = inode;
2367	nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2368	status = ld->prepare_layoutstats(&data->args);
2369	if (status)
2370		goto out_free;
2371
2372	status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2373
2374out:
2375	dprintk("%s returns %d\n", __func__, status);
2376	return status;
2377
2378out_free:
2379	kfree(data);
2380out_put:
2381	pnfs_put_layout_hdr(hdr);
2382	smp_mb__before_atomic();
2383	clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2384	smp_mb__after_atomic();
2385	goto out;
2386}
2387EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2388#endif
2389
2390unsigned int layoutstats_timer;
2391module_param(layoutstats_timer, uint, 0644);
2392EXPORT_SYMBOL_GPL(layoutstats_timer);
2393