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