blob: 2fa483e6dbe2e4a0e4333bf2d76cfd861073232b [file] [log] [blame]
Ricardo Labiaga85e174b2010-10-20 00:17:58 -04001/*
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>
Trond Myklebust493292d2011-07-13 15:58:28 -040031#include <linux/nfs_page.h>
Paul Gortmaker143cb492011-07-01 14:23:34 -040032#include <linux/module.h>
Andy Adamson974cec82010-10-20 00:18:02 -040033#include "internal.h"
Ricardo Labiaga85e174b2010-10-20 00:17:58 -040034#include "pnfs.h"
Andy Adamson64419a92011-03-01 01:34:16 +000035#include "iostat.h"
Trond Myklebustcc668ab2013-08-14 15:31:28 -040036#include "nfs4trace.h"
Trond Myklebust40dd4b72015-01-24 13:54:37 -050037#include "delegation.h"
Peng Tao87334082015-06-23 19:51:57 +080038#include "nfs42.h"
Ricardo Labiaga85e174b2010-10-20 00:17:58 -040039
40#define NFSDBG_FACILITY NFSDBG_PNFS
Trond Myklebust25c75332012-09-18 17:01:12 -040041#define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
Ricardo Labiaga85e174b2010-10-20 00:17:58 -040042
Fred Isaman02c35fc2010-10-20 00:17:59 -040043/* 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
Trond Myklebust13c13a62016-01-26 23:12:11 -050055static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
Peng Taoaa1e0e3a2014-09-06 00:53:25 +080056
Fred Isaman02c35fc2010-10-20 00:17:59 -040057/* Return the registered pnfs layout driver module matching given id */
58static struct pnfs_layoutdriver_type *
59find_pnfs_driver_locked(u32 id)
60{
61 struct pnfs_layoutdriver_type *local;
62
63 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
64 if (local->id == id)
65 goto out;
66 local = NULL;
67out:
68 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
69 return local;
70}
71
Ricardo Labiaga85e174b2010-10-20 00:17:58 -040072static struct pnfs_layoutdriver_type *
73find_pnfs_driver(u32 id)
74{
Fred Isaman02c35fc2010-10-20 00:17:59 -040075 struct pnfs_layoutdriver_type *local;
76
77 spin_lock(&pnfs_spinlock);
78 local = find_pnfs_driver_locked(id);
Trond Myklebust0a9c63f2012-06-15 13:02:58 -040079 if (local != NULL && !try_module_get(local->owner)) {
80 dprintk("%s: Could not grab reference on module\n", __func__);
81 local = NULL;
82 }
Fred Isaman02c35fc2010-10-20 00:17:59 -040083 spin_unlock(&pnfs_spinlock);
84 return local;
Ricardo Labiaga85e174b2010-10-20 00:17:58 -040085}
86
87void
88unset_pnfs_layoutdriver(struct nfs_server *nfss)
89{
Benny Halevy738fd0f32011-07-30 20:52:36 -040090 if (nfss->pnfs_curr_ld) {
91 if (nfss->pnfs_curr_ld->clear_layoutdriver)
92 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
Trond Myklebust2a4c8992012-06-14 13:08:38 -040093 /* Decrement the MDS count. Purge the deviceid cache if zero */
94 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
95 nfs4_deviceid_purge_client(nfss->nfs_client);
Fred Isaman02c35fc2010-10-20 00:17:59 -040096 module_put(nfss->pnfs_curr_ld->owner);
Benny Halevy738fd0f32011-07-30 20:52:36 -040097 }
Ricardo Labiaga85e174b2010-10-20 00:17:58 -040098 nfss->pnfs_curr_ld = NULL;
99}
100
101/*
102 * Try to set the server's pnfs module to the pnfs layout type specified by id.
103 * Currently only one pNFS layout driver per filesystem is supported.
104 *
105 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
106 */
107void
Benny Halevy738fd0f32011-07-30 20:52:36 -0400108set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
109 u32 id)
Ricardo Labiaga85e174b2010-10-20 00:17:58 -0400110{
111 struct pnfs_layoutdriver_type *ld_type = NULL;
112
113 if (id == 0)
114 goto out_no_driver;
115 if (!(server->nfs_client->cl_exchange_flags &
116 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
Weston Andros Adamsona0308892012-01-26 13:32:23 -0500117 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
118 __func__, id, server->nfs_client->cl_exchange_flags);
Ricardo Labiaga85e174b2010-10-20 00:17:58 -0400119 goto out_no_driver;
120 }
121 ld_type = find_pnfs_driver(id);
122 if (!ld_type) {
123 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
124 ld_type = find_pnfs_driver(id);
125 if (!ld_type) {
126 dprintk("%s: No pNFS module found for %u.\n",
127 __func__, id);
128 goto out_no_driver;
129 }
130 }
131 server->pnfs_curr_ld = ld_type;
Benny Halevy738fd0f32011-07-30 20:52:36 -0400132 if (ld_type->set_layoutdriver
133 && ld_type->set_layoutdriver(server, mntfh)) {
Weston Andros Adamsona0308892012-01-26 13:32:23 -0500134 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
135 "driver %u.\n", __func__, id);
Benny Halevy738fd0f32011-07-30 20:52:36 -0400136 module_put(ld_type->owner);
137 goto out_no_driver;
138 }
Trond Myklebust2a4c8992012-06-14 13:08:38 -0400139 /* Bump the MDS count */
140 atomic_inc(&server->nfs_client->cl_mds_count);
Christoph Hellwigea8eecd2011-03-01 01:34:21 +0000141
Ricardo Labiaga85e174b2010-10-20 00:17:58 -0400142 dprintk("%s: pNFS module for %u set\n", __func__, id);
143 return;
144
145out_no_driver:
146 dprintk("%s: Using NFSv4 I/O\n", __func__);
147 server->pnfs_curr_ld = NULL;
148}
Fred Isaman02c35fc2010-10-20 00:17:59 -0400149
150int
151pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
152{
153 int status = -EINVAL;
154 struct pnfs_layoutdriver_type *tmp;
155
156 if (ld_type->id == 0) {
Weston Andros Adamsona0308892012-01-26 13:32:23 -0500157 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
Fred Isaman02c35fc2010-10-20 00:17:59 -0400158 return status;
159 }
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400160 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
Weston Andros Adamsona0308892012-01-26 13:32:23 -0500161 printk(KERN_ERR "NFS: %s Layout driver must provide "
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400162 "alloc_lseg and free_lseg.\n", __func__);
163 return status;
164 }
Fred Isaman02c35fc2010-10-20 00:17:59 -0400165
166 spin_lock(&pnfs_spinlock);
167 tmp = find_pnfs_driver_locked(ld_type->id);
168 if (!tmp) {
169 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
170 status = 0;
171 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
172 ld_type->name);
173 } else {
Weston Andros Adamsona0308892012-01-26 13:32:23 -0500174 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
Fred Isaman02c35fc2010-10-20 00:17:59 -0400175 __func__, ld_type->id);
176 }
177 spin_unlock(&pnfs_spinlock);
178
179 return status;
180}
181EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
182
183void
184pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
185{
186 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
187 spin_lock(&pnfs_spinlock);
188 list_del(&ld_type->pnfs_tblid);
189 spin_unlock(&pnfs_spinlock);
190}
191EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
Benny Halevye5e94012010-10-20 00:18:01 -0400192
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400193/*
194 * pNFS client layout cache
195 */
196
Fred Isamancc6e5342011-01-06 11:36:28 +0000197/* Need to hold i_lock if caller does not already hold reference */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000198void
Trond Myklebust70c3bd22012-09-18 20:51:13 -0400199pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
Benny Halevye5e94012010-10-20 00:18:01 -0400200{
Fred Isamancc6e5342011-01-06 11:36:28 +0000201 atomic_inc(&lo->plh_refcount);
202}
203
Benny Halevy636fb9c2011-05-22 19:51:33 +0300204static struct pnfs_layout_hdr *
205pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
206{
207 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
Trond Myklebust57934272012-09-20 20:37:23 -0400208 return ld->alloc_layout_hdr(ino, gfp_flags);
Benny Halevy636fb9c2011-05-22 19:51:33 +0300209}
210
211static void
212pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
213{
Trond Myklebust9c626382012-09-20 17:31:43 -0400214 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
215 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
216
217 if (!list_empty(&lo->plh_layouts)) {
218 struct nfs_client *clp = server->nfs_client;
219
220 spin_lock(&clp->cl_lock);
221 list_del_init(&lo->plh_layouts);
222 spin_unlock(&clp->cl_lock);
223 }
Peng Tao9fa40752011-07-30 20:52:32 -0400224 put_rpccred(lo->plh_lc_cred);
Trond Myklebust57934272012-09-20 20:37:23 -0400225 return ld->free_layout_hdr(lo);
Benny Halevy636fb9c2011-05-22 19:51:33 +0300226}
227
Fred Isamancc6e5342011-01-06 11:36:28 +0000228static void
Trond Myklebust6622c3e2012-09-20 17:23:11 -0400229pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
Fred Isamancc6e5342011-01-06 11:36:28 +0000230{
Trond Myklebustbb346f62012-09-20 15:52:13 -0400231 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
Fred Isamancc6e5342011-01-06 11:36:28 +0000232 dprintk("%s: freeing layout cache %p\n", __func__, lo);
Trond Myklebustbb346f62012-09-20 15:52:13 -0400233 nfsi->layout = NULL;
234 /* Reset MDS Threshold I/O counters */
235 nfsi->write_io = 0;
236 nfsi->read_io = 0;
Benny Halevye5e94012010-10-20 00:18:01 -0400237}
238
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400239void
Trond Myklebust70c3bd22012-09-18 20:51:13 -0400240pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
Andy Adamson974cec82010-10-20 00:18:02 -0400241{
Fred Isamancc6e5342011-01-06 11:36:28 +0000242 struct inode *inode = lo->plh_inode;
243
Trond Myklebust13c13a62016-01-26 23:12:11 -0500244 pnfs_layoutreturn_before_put_layout_hdr(lo);
245
Fred Isamancc6e5342011-01-06 11:36:28 +0000246 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
Peng Tao566f8732014-10-10 23:25:46 +0800247 if (!list_empty(&lo->plh_segs))
248 WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
Trond Myklebust6622c3e2012-09-20 17:23:11 -0400249 pnfs_detach_layout_hdr(lo);
Fred Isamancc6e5342011-01-06 11:36:28 +0000250 spin_unlock(&inode->i_lock);
Trond Myklebust6622c3e2012-09-20 17:23:11 -0400251 pnfs_free_layout_hdr(lo);
Fred Isamancc6e5342011-01-06 11:36:28 +0000252 }
Andy Adamson974cec82010-10-20 00:18:02 -0400253}
254
Trond Myklebust2454dfe2016-02-22 17:34:59 -0500255/*
256 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
257 *
258 * In order to continue using the pnfs_layout_hdr, a full recovery
259 * is required.
260 * Note that caller must hold inode->i_lock.
261 */
Trond Myklebust9fd4b9f2016-02-22 17:46:34 -0500262static int
Trond Myklebust2454dfe2016-02-22 17:34:59 -0500263pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
264 struct list_head *lseg_list)
265{
266 struct pnfs_layout_range range = {
267 .iomode = IOMODE_ANY,
268 .offset = 0,
269 .length = NFS4_MAX_UINT64,
270 };
271
272 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
Trond Myklebust9fd4b9f2016-02-22 17:46:34 -0500273 return pnfs_mark_matching_lsegs_invalid(lo, lseg_list, &range);
Trond Myklebust2454dfe2016-02-22 17:34:59 -0500274}
275
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400276static int
277pnfs_iomode_to_fail_bit(u32 iomode)
278{
279 return iomode == IOMODE_RW ?
280 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
281}
282
283static void
Trond Myklebust3e621212012-09-24 13:07:16 -0400284pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400285{
Trond Myklebust25c75332012-09-18 17:01:12 -0400286 lo->plh_retry_timestamp = jiffies;
Yanchuan Nian39e88fc2013-01-04 20:19:49 +0800287 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
Trond Myklebust3e621212012-09-24 13:07:16 -0400288 atomic_inc(&lo->plh_refcount);
289}
290
291static void
292pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
293{
294 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
295 atomic_dec(&lo->plh_refcount);
296}
297
298static void
299pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
300{
301 struct inode *inode = lo->plh_inode;
Trond Myklebust115ce572012-09-20 21:19:43 -0400302 struct pnfs_layout_range range = {
303 .iomode = iomode,
304 .offset = 0,
305 .length = NFS4_MAX_UINT64,
306 };
307 LIST_HEAD(head);
Trond Myklebust3e621212012-09-24 13:07:16 -0400308
309 spin_lock(&inode->i_lock);
310 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
Trond Myklebust115ce572012-09-20 21:19:43 -0400311 pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
Trond Myklebust3e621212012-09-24 13:07:16 -0400312 spin_unlock(&inode->i_lock);
Trond Myklebust115ce572012-09-20 21:19:43 -0400313 pnfs_free_lseg_list(&head);
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400314 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
315 iomode == IOMODE_RW ? "RW" : "READ");
316}
317
318static bool
319pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
320{
Trond Myklebust25c75332012-09-18 17:01:12 -0400321 unsigned long start, end;
Trond Myklebust3e621212012-09-24 13:07:16 -0400322 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
323
324 if (test_bit(fail_bit, &lo->plh_flags) == 0)
Trond Myklebust25c75332012-09-18 17:01:12 -0400325 return false;
326 end = jiffies;
327 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
328 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
329 /* It is time to retry the failed layoutgets */
Trond Myklebust3e621212012-09-24 13:07:16 -0400330 pnfs_layout_clear_fail_bit(lo, fail_bit);
Trond Myklebust25c75332012-09-18 17:01:12 -0400331 return false;
332 }
333 return true;
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400334}
335
Andy Adamson974cec82010-10-20 00:18:02 -0400336static void
337init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
338{
Fred Isaman566052c2011-01-06 11:36:20 +0000339 INIT_LIST_HEAD(&lseg->pls_list);
Peng Taoa9bae562011-07-30 20:52:33 -0400340 INIT_LIST_HEAD(&lseg->pls_lc_list);
Fred Isaman4541d162011-01-06 11:36:23 +0000341 atomic_set(&lseg->pls_refcount, 1);
342 smp_mb();
343 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
Fred Isaman566052c2011-01-06 11:36:20 +0000344 lseg->pls_layout = lo;
Andy Adamson974cec82010-10-20 00:18:02 -0400345}
346
Trond Myklebust905ca192012-09-20 20:46:49 -0400347static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
Andy Adamson974cec82010-10-20 00:18:02 -0400348{
Fred Isamanb7edfaa2011-01-06 11:36:21 +0000349 struct inode *ino = lseg->pls_layout->plh_inode;
Andy Adamson974cec82010-10-20 00:18:02 -0400350
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400351 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400352}
353
Fred Isamand684d2a2011-03-01 01:34:13 +0000354static void
Trond Myklebust57036a32012-09-20 16:33:30 -0400355pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
356 struct pnfs_layout_segment *lseg)
Andy Adamson974cec82010-10-20 00:18:02 -0400357{
Trond Myklebust57036a32012-09-20 16:33:30 -0400358 struct inode *inode = lo->plh_inode;
Fred Isamand684d2a2011-03-01 01:34:13 +0000359
Benny Halevyd20581a2011-05-22 19:52:03 +0300360 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
Fred Isamand684d2a2011-03-01 01:34:13 +0000361 list_del_init(&lseg->pls_list);
Trond Myklebust8f0d27d2012-09-20 20:57:11 -0400362 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
363 atomic_dec(&lo->plh_refcount);
Trond Myklebust173f77e9c2012-09-21 15:49:42 -0400364 if (list_empty(&lo->plh_segs))
365 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
Fred Isamand684d2a2011-03-01 01:34:13 +0000366 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
367}
368
Fred Isamanbae724e2011-03-01 01:34:15 +0000369void
Trond Myklebust9369a432012-09-18 20:57:08 -0400370pnfs_put_lseg(struct pnfs_layout_segment *lseg)
Fred Isamand684d2a2011-03-01 01:34:13 +0000371{
Trond Myklebust57036a32012-09-20 16:33:30 -0400372 struct pnfs_layout_hdr *lo;
Fred Isamand684d2a2011-03-01 01:34:13 +0000373 struct inode *inode;
374
375 if (!lseg)
376 return;
377
Fred Isaman4541d162011-01-06 11:36:23 +0000378 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
379 atomic_read(&lseg->pls_refcount),
380 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
Trond Myklebust4ef2e4f2015-02-05 17:27:39 -0500381
Trond Myklebust57036a32012-09-20 16:33:30 -0400382 lo = lseg->pls_layout;
383 inode = lo->plh_inode;
Trond Myklebust4ef2e4f2015-02-05 17:27:39 -0500384
Fred Isamand684d2a2011-03-01 01:34:13 +0000385 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
Trond Myklebustfaa4a542015-07-09 18:24:07 +0200386 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
387 spin_unlock(&inode->i_lock);
388 return;
389 }
Trond Myklebust8f0d27d2012-09-20 20:57:11 -0400390 pnfs_get_layout_hdr(lo);
Trond Myklebust4ef2e4f2015-02-05 17:27:39 -0500391 pnfs_layout_remove_lseg(lo, lseg);
392 spin_unlock(&inode->i_lock);
393 pnfs_free_lseg(lseg);
394 pnfs_put_layout_hdr(lo);
Fred Isaman4541d162011-01-06 11:36:23 +0000395 }
Andy Adamson974cec82010-10-20 00:18:02 -0400396}
Trond Myklebust9369a432012-09-18 20:57:08 -0400397EXPORT_SYMBOL_GPL(pnfs_put_lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400398
Trond Myklebust6543f802014-10-08 16:39:12 -0400399static void pnfs_free_lseg_async_work(struct work_struct *work)
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400400{
401 struct pnfs_layout_segment *lseg;
Trond Myklebust6543f802014-10-08 16:39:12 -0400402 struct pnfs_layout_hdr *lo;
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400403
404 lseg = container_of(work, struct pnfs_layout_segment, pls_work);
Trond Myklebust6543f802014-10-08 16:39:12 -0400405 lo = lseg->pls_layout;
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400406
Trond Myklebust6543f802014-10-08 16:39:12 -0400407 pnfs_free_lseg(lseg);
408 pnfs_put_layout_hdr(lo);
409}
410
411static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
412{
413 INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
414 schedule_work(&lseg->pls_work);
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400415}
416
417void
Trond Myklebust6543f802014-10-08 16:39:12 -0400418pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400419{
Trond Myklebust6543f802014-10-08 16:39:12 -0400420 if (!lseg)
421 return;
422
423 assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
424
425 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
426 atomic_read(&lseg->pls_refcount),
427 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
428 if (atomic_dec_and_test(&lseg->pls_refcount)) {
429 struct pnfs_layout_hdr *lo = lseg->pls_layout;
Trond Myklebustfaa4a542015-07-09 18:24:07 +0200430 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
431 return;
Trond Myklebust6543f802014-10-08 16:39:12 -0400432 pnfs_get_layout_hdr(lo);
433 pnfs_layout_remove_lseg(lo, lseg);
434 pnfs_free_lseg_async(lseg);
435 }
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400436}
Trond Myklebust6543f802014-10-08 16:39:12 -0400437EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400438
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400439static u64
Benny Halevyfb3296e2011-05-22 19:47:26 +0300440end_offset(u64 start, u64 len)
Fred Isaman4541d162011-01-06 11:36:23 +0000441{
Benny Halevyfb3296e2011-05-22 19:47:26 +0300442 u64 end;
443
444 end = start + len;
445 return end >= start ? end : NFS4_MAX_UINT64;
446}
447
Benny Halevyfb3296e2011-05-22 19:47:26 +0300448/*
449 * is l2 fully contained in l1?
450 * start1 end1
451 * [----------------------------------)
452 * start2 end2
453 * [----------------)
454 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400455static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400456pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400457 const struct pnfs_layout_range *l2)
Benny Halevyfb3296e2011-05-22 19:47:26 +0300458{
459 u64 start1 = l1->offset;
460 u64 end1 = end_offset(start1, l1->length);
461 u64 start2 = l2->offset;
462 u64 end2 = end_offset(start2, l2->length);
463
464 return (start1 <= start2) && (end1 >= end2);
465}
466
467/*
468 * is l1 and l2 intersecting?
469 * start1 end1
470 * [----------------------------------)
471 * start2 end2
472 * [----------------)
473 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400474static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400475pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400476 const struct pnfs_layout_range *l2)
Benny Halevyfb3296e2011-05-22 19:47:26 +0300477{
478 u64 start1 = l1->offset;
479 u64 end1 = end_offset(start1, l1->length);
480 u64 start2 = l2->offset;
481 u64 end2 = end_offset(start2, l2->length);
482
483 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
484 (end2 == NFS4_MAX_UINT64 || end2 > start1);
485}
486
Fred Isaman4541d162011-01-06 11:36:23 +0000487static bool
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400488should_free_lseg(const struct pnfs_layout_range *lseg_range,
489 const struct pnfs_layout_range *recall_range)
Fred Isaman4541d162011-01-06 11:36:23 +0000490{
Benny Halevy778b5502011-05-22 19:48:02 +0300491 return (recall_range->iomode == IOMODE_ANY ||
492 lseg_range->iomode == recall_range->iomode) &&
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400493 pnfs_lseg_range_intersecting(lseg_range, recall_range);
Fred Isaman4541d162011-01-06 11:36:23 +0000494}
495
Trond Myklebust24956802013-03-20 13:03:00 -0400496static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
497 struct list_head *tmp_list)
498{
499 if (!atomic_dec_and_test(&lseg->pls_refcount))
500 return false;
501 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
502 list_add(&lseg->pls_list, tmp_list);
503 return true;
504}
505
Fred Isaman4541d162011-01-06 11:36:23 +0000506/* Returns 1 if lseg is removed from list, 0 otherwise */
507static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
508 struct list_head *tmp_list)
509{
510 int rv = 0;
511
512 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
513 /* Remove the reference keeping the lseg in the
514 * list. It will now be removed when all
515 * outstanding io is finished.
516 */
Fred Isamand684d2a2011-03-01 01:34:13 +0000517 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
518 atomic_read(&lseg->pls_refcount));
Trond Myklebust24956802013-03-20 13:03:00 -0400519 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
Fred Isamand684d2a2011-03-01 01:34:13 +0000520 rv = 1;
Fred Isaman4541d162011-01-06 11:36:23 +0000521 }
522 return rv;
523}
524
525/* Returns count of number of matching invalid lsegs remaining in list
526 * after call.
527 */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000528int
Trond Myklebust49a85062012-09-18 20:43:31 -0400529pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
Fred Isaman4541d162011-01-06 11:36:23 +0000530 struct list_head *tmp_list,
Trond Myklebust506c0d62016-01-04 13:04:47 -0500531 const struct pnfs_layout_range *recall_range)
Andy Adamson974cec82010-10-20 00:18:02 -0400532{
533 struct pnfs_layout_segment *lseg, *next;
Trond Myklebust71b39852016-01-04 12:41:15 -0500534 int remaining = 0;
Andy Adamson974cec82010-10-20 00:18:02 -0400535
536 dprintk("%s:Begin lo %p\n", __func__, lo);
537
Trond Myklebust8006bfb2012-09-21 14:48:04 -0400538 if (list_empty(&lo->plh_segs))
Fred Isaman385117222011-02-03 18:28:50 +0000539 return 0;
Fred Isaman4541d162011-01-06 11:36:23 +0000540 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
Benny Halevy778b5502011-05-22 19:48:02 +0300541 if (!recall_range ||
542 should_free_lseg(&lseg->pls_range, recall_range)) {
Fred Isaman4541d162011-01-06 11:36:23 +0000543 dprintk("%s: freeing lseg %p iomode %d "
544 "offset %llu length %llu\n", __func__,
545 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
546 lseg->pls_range.length);
Trond Myklebust71b39852016-01-04 12:41:15 -0500547 if (!mark_lseg_invalid(lseg, tmp_list))
548 remaining++;
Fred Isaman4541d162011-01-06 11:36:23 +0000549 }
Trond Myklebust71b39852016-01-04 12:41:15 -0500550 dprintk("%s:Return %i\n", __func__, remaining);
551 return remaining;
Andy Adamson974cec82010-10-20 00:18:02 -0400552}
553
Fred Isamanf49f9ba2011-02-03 18:28:52 +0000554/* note free_me must contain lsegs from a single layout_hdr */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000555void
Fred Isaman4541d162011-01-06 11:36:23 +0000556pnfs_free_lseg_list(struct list_head *free_me)
Andy Adamson974cec82010-10-20 00:18:02 -0400557{
Fred Isaman4541d162011-01-06 11:36:23 +0000558 struct pnfs_layout_segment *lseg, *tmp;
Andy Adamson974cec82010-10-20 00:18:02 -0400559
Fred Isamanf49f9ba2011-02-03 18:28:52 +0000560 if (list_empty(free_me))
561 return;
562
Fred Isaman4541d162011-01-06 11:36:23 +0000563 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
Fred Isaman566052c2011-01-06 11:36:20 +0000564 list_del(&lseg->pls_list);
Trond Myklebust905ca192012-09-20 20:46:49 -0400565 pnfs_free_lseg(lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400566 }
567}
568
Benny Halevye5e94012010-10-20 00:18:01 -0400569void
570pnfs_destroy_layout(struct nfs_inode *nfsi)
571{
572 struct pnfs_layout_hdr *lo;
Andy Adamson974cec82010-10-20 00:18:02 -0400573 LIST_HEAD(tmp_list);
Benny Halevye5e94012010-10-20 00:18:01 -0400574
575 spin_lock(&nfsi->vfs_inode.i_lock);
576 lo = nfsi->layout;
577 if (lo) {
Trond Myklebust3e621212012-09-24 13:07:16 -0400578 pnfs_get_layout_hdr(lo);
Trond Myklebust2454dfe2016-02-22 17:34:59 -0500579 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
Trond Myklebust3e621212012-09-24 13:07:16 -0400580 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
581 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
582 spin_unlock(&nfsi->vfs_inode.i_lock);
583 pnfs_free_lseg_list(&tmp_list);
584 pnfs_put_layout_hdr(lo);
585 } else
586 spin_unlock(&nfsi->vfs_inode.i_lock);
Benny Halevye5e94012010-10-20 00:18:01 -0400587}
Andy Adamson041245c2012-04-27 17:53:53 -0400588EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
Benny Halevye5e94012010-10-20 00:18:01 -0400589
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500590static bool
591pnfs_layout_add_bulk_destroy_list(struct inode *inode,
592 struct list_head *layout_list)
593{
594 struct pnfs_layout_hdr *lo;
595 bool ret = false;
596
597 spin_lock(&inode->i_lock);
598 lo = NFS_I(inode)->layout;
599 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
600 pnfs_get_layout_hdr(lo);
601 list_add(&lo->plh_bulk_destroy, layout_list);
602 ret = true;
603 }
604 spin_unlock(&inode->i_lock);
605 return ret;
606}
607
608/* Caller must hold rcu_read_lock and clp->cl_lock */
609static int
610pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
611 struct nfs_server *server,
612 struct list_head *layout_list)
613{
614 struct pnfs_layout_hdr *lo, *next;
615 struct inode *inode;
616
617 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
618 inode = igrab(lo->plh_inode);
619 if (inode == NULL)
620 continue;
621 list_del_init(&lo->plh_layouts);
622 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
623 continue;
624 rcu_read_unlock();
625 spin_unlock(&clp->cl_lock);
626 iput(inode);
627 spin_lock(&clp->cl_lock);
628 rcu_read_lock();
629 return -EAGAIN;
630 }
631 return 0;
632}
633
634static int
635pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
636 bool is_bulk_recall)
637{
638 struct pnfs_layout_hdr *lo;
639 struct inode *inode;
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500640 LIST_HEAD(lseg_list);
641 int ret = 0;
642
643 while (!list_empty(layout_list)) {
644 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
645 plh_bulk_destroy);
646 dprintk("%s freeing layout for inode %lu\n", __func__,
647 lo->plh_inode->i_ino);
648 inode = lo->plh_inode;
Christoph Hellwig7c5d1872014-09-10 08:23:29 -0700649
650 pnfs_layoutcommit_inode(inode, false);
651
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500652 spin_lock(&inode->i_lock);
653 list_del_init(&lo->plh_bulk_destroy);
Trond Myklebust9fd4b9f2016-02-22 17:46:34 -0500654 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
655 if (is_bulk_recall)
656 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500657 ret = -EAGAIN;
Trond Myklebust9fd4b9f2016-02-22 17:46:34 -0500658 }
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500659 spin_unlock(&inode->i_lock);
660 pnfs_free_lseg_list(&lseg_list);
Trond Myklebustb20135d2015-12-31 09:28:06 -0500661 /* Free all lsegs that are attached to commit buckets */
662 nfs_commit_inode(inode, 0);
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500663 pnfs_put_layout_hdr(lo);
664 iput(inode);
665 }
666 return ret;
667}
668
669int
670pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
671 struct nfs_fsid *fsid,
672 bool is_recall)
673{
674 struct nfs_server *server;
675 LIST_HEAD(layout_list);
676
677 spin_lock(&clp->cl_lock);
678 rcu_read_lock();
679restart:
680 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
681 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
682 continue;
683 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
684 server,
685 &layout_list) != 0)
686 goto restart;
687 }
688 rcu_read_unlock();
689 spin_unlock(&clp->cl_lock);
690
691 if (list_empty(&layout_list))
692 return 0;
693 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
694}
695
696int
697pnfs_destroy_layouts_byclid(struct nfs_client *clp,
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 (pnfs_layout_bulk_destroy_byserver_locked(clp,
708 server,
709 &layout_list) != 0)
710 goto restart;
711 }
712 rcu_read_unlock();
713 spin_unlock(&clp->cl_lock);
714
715 if (list_empty(&layout_list))
716 return 0;
717 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
718}
719
Andy Adamson974cec82010-10-20 00:18:02 -0400720/*
721 * Called by the state manger to remove all layouts established under an
722 * expired lease.
723 */
724void
725pnfs_destroy_all_layouts(struct nfs_client *clp)
726{
Andy Adamsonc47abcf2011-06-15 17:52:40 -0400727 nfs4_deviceid_mark_client_invalid(clp);
728 nfs4_deviceid_purge_client(clp);
729
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500730 pnfs_destroy_layouts_byclid(clp, false);
Andy Adamson974cec82010-10-20 00:18:02 -0400731}
732
Trond Myklebust5a655032012-10-02 16:47:14 -0700733/*
734 * Compare 2 layout stateid sequence ids, to see which is newer,
735 * taking into account wraparound issues.
736 */
737static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
738{
Trond Myklebust2c64c572014-02-12 09:36:59 -0500739 return (s32)(s1 - s2) > 0;
Trond Myklebust5a655032012-10-02 16:47:14 -0700740}
741
Fred Isamanfd6002e2011-01-06 11:36:22 +0000742/* update lo->plh_stateid with new if is more recent */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000743void
744pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
745 bool update_barrier)
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400746{
Trond Myklebust22aaf712012-10-04 16:32:22 -0700747 u32 oldseq, newseq, new_barrier;
748 int empty = list_empty(&lo->plh_segs);
Andy Adamson974cec82010-10-20 00:18:02 -0400749
Trond Myklebust2d2f24a2012-03-04 18:13:57 -0500750 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
751 newseq = be32_to_cpu(new->seqid);
Trond Myklebust22aaf712012-10-04 16:32:22 -0700752 if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
Trond Myklebustf597c532012-03-04 18:13:56 -0500753 nfs4_stateid_copy(&lo->plh_stateid, new);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000754 if (update_barrier) {
Trond Myklebust22aaf712012-10-04 16:32:22 -0700755 new_barrier = be32_to_cpu(new->seqid);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000756 } else {
757 /* Because of wraparound, we want to keep the barrier
Trond Myklebust22aaf712012-10-04 16:32:22 -0700758 * "close" to the current seqids.
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000759 */
Trond Myklebust22aaf712012-10-04 16:32:22 -0700760 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000761 }
Trond Myklebust22aaf712012-10-04 16:32:22 -0700762 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
763 lo->plh_barrier = new_barrier;
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000764 }
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400765}
766
Trond Myklebust19c54ab2012-10-05 16:56:58 -0700767static bool
768pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
769 const nfs4_stateid *stateid)
770{
771 u32 seqid = be32_to_cpu(stateid->seqid);
772
773 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
774}
775
Fred Isamancf7d63f2011-01-06 11:36:25 +0000776/* lget is set to 1 if called from inside send_layoutget call chain */
777static bool
Trond Myklebuste1c06f82015-08-04 16:40:08 -0400778pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
Fred Isamancf7d63f2011-01-06 11:36:25 +0000779{
Fred Isamanf7e89172011-01-06 11:36:32 +0000780 return lo->plh_block_lgets ||
Trond Myklebuste1c06f82015-08-04 16:40:08 -0400781 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
Fred Isamancf7d63f2011-01-06 11:36:25 +0000782}
783
Fred Isamanfd6002e2011-01-06 11:36:22 +0000784int
785pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
Trond Myklebust506c0d62016-01-04 13:04:47 -0500786 const struct pnfs_layout_range *range,
Fred Isamanfd6002e2011-01-06 11:36:22 +0000787 struct nfs4_state *open_state)
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400788{
Fred Isamanfd6002e2011-01-06 11:36:22 +0000789 int status = 0;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400790
791 dprintk("--> %s\n", __func__);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000792 spin_lock(&lo->plh_inode->i_lock);
Trond Myklebuste1c06f82015-08-04 16:40:08 -0400793 if (pnfs_layoutgets_blocked(lo)) {
Fred Isamancf7d63f2011-01-06 11:36:25 +0000794 status = -EAGAIN;
Trond Myklebust5d422302013-03-14 16:57:48 -0400795 } else if (!nfs4_valid_open_stateid(open_state)) {
796 status = -EBADF;
Christoph Hellwig47abade2014-08-21 11:09:22 -0500797 } else if (list_empty(&lo->plh_segs) ||
798 test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
Fred Isamanfd6002e2011-01-06 11:36:22 +0000799 int seq;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400800
Fred Isamanfd6002e2011-01-06 11:36:22 +0000801 do {
802 seq = read_seqbegin(&open_state->seqlock);
Trond Myklebustf597c532012-03-04 18:13:56 -0500803 nfs4_stateid_copy(dst, &open_state->stateid);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000804 } while (read_seqretry(&open_state->seqlock, seq));
805 } else
Trond Myklebustf597c532012-03-04 18:13:56 -0500806 nfs4_stateid_copy(dst, &lo->plh_stateid);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000807 spin_unlock(&lo->plh_inode->i_lock);
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400808 dprintk("<-- %s\n", __func__);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000809 return status;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400810}
811
812/*
813* Get layout from server.
814* for now, assume that whole file layouts are requested.
815* arg->offset: 0
816* arg->length: all ones
817*/
Benny Halevye5e94012010-10-20 00:18:01 -0400818static struct pnfs_layout_segment *
819send_layoutget(struct pnfs_layout_hdr *lo,
820 struct nfs_open_context *ctx,
Trond Myklebuste144e532016-01-04 12:52:53 -0500821 const struct pnfs_layout_range *range,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400822 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -0400823{
Fred Isamanb7edfaa2011-01-06 11:36:21 +0000824 struct inode *ino = lo->plh_inode;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400825 struct nfs_server *server = NFS_SERVER(ino);
826 struct nfs4_layoutget *lgp;
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -0400827 struct pnfs_layout_segment *lseg;
Trond Myklebust2d89a1d2015-08-31 02:05:47 -0700828 loff_t i_size;
Benny Halevye5e94012010-10-20 00:18:01 -0400829
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400830 dprintk("--> %s\n", __func__);
831
Jeff Layton4f2e9dc2015-11-25 13:43:14 -0500832 /*
833 * Synchronously retrieve layout information from server and
834 * store in lseg. If we race with a concurrent seqid morphing
835 * op, then re-send the LAYOUTGET.
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400836 */
Jeff Layton4f2e9dc2015-11-25 13:43:14 -0500837 do {
838 lgp = kzalloc(sizeof(*lgp), gfp_flags);
839 if (lgp == NULL)
840 return NULL;
841
842 i_size = i_size_read(ino);
843
844 lgp->args.minlength = PAGE_CACHE_SIZE;
845 if (lgp->args.minlength > range->length)
846 lgp->args.minlength = range->length;
847 if (range->iomode == IOMODE_READ) {
848 if (range->offset >= i_size)
849 lgp->args.minlength = 0;
850 else if (i_size - range->offset < lgp->args.minlength)
851 lgp->args.minlength = i_size - range->offset;
852 }
853 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
Trond Myklebuste144e532016-01-04 12:52:53 -0500854 pnfs_copy_range(&lgp->args.range, range);
Jeff Layton4f2e9dc2015-11-25 13:43:14 -0500855 lgp->args.type = server->pnfs_curr_ld->id;
856 lgp->args.inode = ino;
857 lgp->args.ctx = get_nfs_open_context(ctx);
858 lgp->gfp_flags = gfp_flags;
859 lgp->cred = lo->plh_lc_cred;
860
861 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
862 } while (lseg == ERR_PTR(-EAGAIN));
863
Peng Tao0bcbf032015-12-05 15:57:31 +0800864 if (IS_ERR(lseg) && !nfs_error_is_fatal(PTR_ERR(lseg)))
865 lseg = NULL;
866 else
Tom Haynesd67ae822014-12-11 17:02:04 -0500867 pnfs_layout_clear_fail_bit(lo,
868 pnfs_iomode_to_fail_bit(range->iomode));
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400869
Andy Adamson974cec82010-10-20 00:18:02 -0400870 return lseg;
871}
872
Trond Myklebust24956802013-03-20 13:03:00 -0400873static void pnfs_clear_layoutcommit(struct inode *inode,
874 struct list_head *head)
875{
876 struct nfs_inode *nfsi = NFS_I(inode);
877 struct pnfs_layout_segment *lseg, *tmp;
878
879 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
880 return;
881 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
882 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
883 continue;
884 pnfs_lseg_dec_and_remove_zero(lseg, head);
885 }
886}
887
Tom Haynesd67ae822014-12-11 17:02:04 -0500888void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
889{
890 clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
891 smp_mb__after_atomic();
892 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
Trond Myklebust7f273922015-07-09 18:40:01 +0200893 rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
Tom Haynesd67ae822014-12-11 17:02:04 -0500894}
895
Trond Myklebust13c13a62016-01-26 23:12:11 -0500896static bool
897pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
898{
899 if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
900 return false;
901 lo->plh_return_iomode = 0;
902 pnfs_get_layout_hdr(lo);
Trond Myklebust2370abd2016-01-27 20:32:50 -0500903 clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
Trond Myklebust13c13a62016-01-26 23:12:11 -0500904 return true;
905}
906
Peng Taof40eb5d2014-09-06 00:53:22 +0800907static int
Trond Myklebusted429d62016-01-04 12:30:55 -0500908pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
Peng Tao6c166052014-11-17 09:30:40 +0800909 enum pnfs_iomode iomode, bool sync)
Peng Taof40eb5d2014-09-06 00:53:22 +0800910{
911 struct inode *ino = lo->plh_inode;
912 struct nfs4_layoutreturn *lrp;
913 int status = 0;
914
Trond Myklebuste4af4402015-02-05 17:05:08 -0500915 lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
Peng Taof40eb5d2014-09-06 00:53:22 +0800916 if (unlikely(lrp == NULL)) {
917 status = -ENOMEM;
918 spin_lock(&ino->i_lock);
Tom Haynesd67ae822014-12-11 17:02:04 -0500919 pnfs_clear_layoutreturn_waitbit(lo);
Peng Taof40eb5d2014-09-06 00:53:22 +0800920 spin_unlock(&ino->i_lock);
921 pnfs_put_layout_hdr(lo);
922 goto out;
923 }
924
Trond Myklebusted429d62016-01-04 12:30:55 -0500925 nfs4_stateid_copy(&lrp->args.stateid, stateid);
Peng Taof40eb5d2014-09-06 00:53:22 +0800926 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
927 lrp->args.inode = ino;
Peng Tao15eb67c2014-11-17 09:30:36 +0800928 lrp->args.range.iomode = iomode;
929 lrp->args.range.offset = 0;
930 lrp->args.range.length = NFS4_MAX_UINT64;
Peng Taof40eb5d2014-09-06 00:53:22 +0800931 lrp->args.layout = lo;
932 lrp->clp = NFS_SERVER(ino)->nfs_client;
933 lrp->cred = lo->plh_lc_cred;
934
Peng Tao6c166052014-11-17 09:30:40 +0800935 status = nfs4_proc_layoutreturn(lrp, sync);
Peng Taof40eb5d2014-09-06 00:53:22 +0800936out:
937 dprintk("<-- %s status: %d\n", __func__, status);
938 return status;
939}
940
Trond Myklebust13c13a62016-01-26 23:12:11 -0500941/* Return true if layoutreturn is needed */
942static bool
943pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
944{
945 struct pnfs_layout_segment *s;
946
Trond Myklebust2370abd2016-01-27 20:32:50 -0500947 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
Trond Myklebust13c13a62016-01-26 23:12:11 -0500948 return false;
949
950 /* Defer layoutreturn until all lsegs are done */
951 list_for_each_entry(s, &lo->plh_segs, pls_list) {
952 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
953 return false;
954 }
955
956 return true;
957}
958
959static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
960{
961 struct inode *inode= lo->plh_inode;
962
Trond Myklebust2370abd2016-01-27 20:32:50 -0500963 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
Trond Myklebust13c13a62016-01-26 23:12:11 -0500964 return;
965 spin_lock(&inode->i_lock);
966 if (pnfs_layout_need_return(lo)) {
967 nfs4_stateid stateid;
968 enum pnfs_iomode iomode;
969 bool send;
970
971 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
972 iomode = lo->plh_return_iomode;
973 send = pnfs_prepare_layoutreturn(lo);
974 spin_unlock(&inode->i_lock);
975 if (send) {
976 /* Send an async layoutreturn so we dont deadlock */
977 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
978 }
979 } else
980 spin_unlock(&inode->i_lock);
981}
982
Andy Adamson293b3b02012-06-20 15:03:34 -0400983/*
984 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
985 * when the layout segment list is empty.
986 *
987 * Note that a pnfs_layout_hdr can exist with an empty layout segment
988 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
989 * deviceid is marked invalid.
990 */
Benny Halevycbe82602011-05-22 19:52:37 +0300991int
992_pnfs_return_layout(struct inode *ino)
993{
994 struct pnfs_layout_hdr *lo = NULL;
995 struct nfs_inode *nfsi = NFS_I(ino);
996 LIST_HEAD(tmp_list);
Benny Halevycbe82602011-05-22 19:52:37 +0300997 nfs4_stateid stateid;
Andy Adamson293b3b02012-06-20 15:03:34 -0400998 int status = 0, empty;
Trond Myklebust7f273922015-07-09 18:40:01 +0200999 bool send;
Benny Halevycbe82602011-05-22 19:52:37 +03001000
Andy Adamson366d5052012-06-20 15:03:33 -04001001 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
Benny Halevycbe82602011-05-22 19:52:37 +03001002
1003 spin_lock(&ino->i_lock);
1004 lo = nfsi->layout;
Trond Myklebuste5929f32012-09-21 16:37:02 -04001005 if (!lo) {
Benny Halevycbe82602011-05-22 19:52:37 +03001006 spin_unlock(&ino->i_lock);
Andy Adamson293b3b02012-06-20 15:03:34 -04001007 dprintk("NFS: %s no layout to return\n", __func__);
1008 goto out;
Benny Halevycbe82602011-05-22 19:52:37 +03001009 }
Trond Myklebust50f563e2016-01-04 12:22:46 -05001010 nfs4_stateid_copy(&stateid, &nfsi->layout->plh_stateid);
Benny Halevycbe82602011-05-22 19:52:37 +03001011 /* Reference matched in nfs4_layoutreturn_release */
Trond Myklebust70c3bd22012-09-18 20:51:13 -04001012 pnfs_get_layout_hdr(lo);
Andy Adamson293b3b02012-06-20 15:03:34 -04001013 empty = list_empty(&lo->plh_segs);
Trond Myklebust24956802013-03-20 13:03:00 -04001014 pnfs_clear_layoutcommit(ino, &tmp_list);
Trond Myklebust49a85062012-09-18 20:43:31 -04001015 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
Christoph Hellwigc88953d2014-09-10 08:23:31 -07001016
1017 if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1018 struct pnfs_layout_range range = {
1019 .iomode = IOMODE_ANY,
1020 .offset = 0,
1021 .length = NFS4_MAX_UINT64,
1022 };
1023 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1024 }
1025
Andy Adamson293b3b02012-06-20 15:03:34 -04001026 /* Don't send a LAYOUTRETURN if list was initially empty */
1027 if (empty) {
1028 spin_unlock(&ino->i_lock);
Andy Adamson293b3b02012-06-20 15:03:34 -04001029 dprintk("NFS: %s no layout segments to return\n", __func__);
Trond Myklebust7f273922015-07-09 18:40:01 +02001030 goto out_put_layout_hdr;
Andy Adamson293b3b02012-06-20 15:03:34 -04001031 }
Christoph Hellwig47abade2014-08-21 11:09:22 -05001032
1033 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
Trond Myklebust7f273922015-07-09 18:40:01 +02001034 send = pnfs_prepare_layoutreturn(lo);
Benny Halevycbe82602011-05-22 19:52:37 +03001035 spin_unlock(&ino->i_lock);
1036 pnfs_free_lseg_list(&tmp_list);
Trond Myklebust7f273922015-07-09 18:40:01 +02001037 if (send)
Trond Myklebusted429d62016-01-04 12:30:55 -05001038 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
Trond Myklebust7f273922015-07-09 18:40:01 +02001039out_put_layout_hdr:
1040 pnfs_put_layout_hdr(lo);
Benny Halevycbe82602011-05-22 19:52:37 +03001041out:
1042 dprintk("<-- %s status: %d\n", __func__, status);
1043 return status;
1044}
Andy Adamson0a57cda2012-04-27 17:53:50 -04001045EXPORT_SYMBOL_GPL(_pnfs_return_layout);
Benny Halevycbe82602011-05-22 19:52:37 +03001046
Trond Myklebust24028672013-03-20 13:23:33 -04001047int
1048pnfs_commit_and_return_layout(struct inode *inode)
1049{
1050 struct pnfs_layout_hdr *lo;
1051 int ret;
1052
1053 spin_lock(&inode->i_lock);
1054 lo = NFS_I(inode)->layout;
1055 if (lo == NULL) {
1056 spin_unlock(&inode->i_lock);
1057 return 0;
1058 }
1059 pnfs_get_layout_hdr(lo);
1060 /* Block new layoutgets and read/write to ds */
1061 lo->plh_block_lgets++;
1062 spin_unlock(&inode->i_lock);
1063 filemap_fdatawait(inode->i_mapping);
1064 ret = pnfs_layoutcommit_inode(inode, true);
1065 if (ret == 0)
1066 ret = _pnfs_return_layout(inode);
1067 spin_lock(&inode->i_lock);
1068 lo->plh_block_lgets--;
1069 spin_unlock(&inode->i_lock);
1070 pnfs_put_layout_hdr(lo);
1071 return ret;
1072}
1073
Fred Isamanf7e89172011-01-06 11:36:32 +00001074bool pnfs_roc(struct inode *ino)
1075{
Trond Myklebust40dd4b72015-01-24 13:54:37 -05001076 struct nfs_inode *nfsi = NFS_I(ino);
1077 struct nfs_open_context *ctx;
1078 struct nfs4_state *state;
Fred Isamanf7e89172011-01-06 11:36:32 +00001079 struct pnfs_layout_hdr *lo;
1080 struct pnfs_layout_segment *lseg, *tmp;
Peng Tao193e3aa2014-11-17 09:30:41 +08001081 nfs4_stateid stateid;
Fred Isamanf7e89172011-01-06 11:36:32 +00001082 LIST_HEAD(tmp_list);
Peng Taoe755d632015-08-19 13:49:19 +08001083 bool found = false, layoutreturn = false, roc = false;
Fred Isamanf7e89172011-01-06 11:36:32 +00001084
1085 spin_lock(&ino->i_lock);
Trond Myklebust40dd4b72015-01-24 13:54:37 -05001086 lo = nfsi->layout;
Peng Tao39761432015-08-21 12:49:44 +08001087 if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
Trond Myklebust40dd4b72015-01-24 13:54:37 -05001088 goto out_noroc;
1089
Peng Taoe755d632015-08-19 13:49:19 +08001090 /* no roc if we hold a delegation */
Trond Myklebust40dd4b72015-01-24 13:54:37 -05001091 if (nfs4_check_delegation(ino, FMODE_READ))
1092 goto out_noroc;
1093
1094 list_for_each_entry(ctx, &nfsi->open_files, list) {
1095 state = ctx->state;
1096 /* Don't return layout if there is open file state */
1097 if (state != NULL && state->state != 0)
1098 goto out_noroc;
1099 }
1100
Trond Myklebust50f563e2016-01-04 12:22:46 -05001101 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
Peng Taoe755d632015-08-19 13:49:19 +08001102 /* always send layoutreturn if being marked so */
Trond Myklebust2370abd2016-01-27 20:32:50 -05001103 if (test_and_clear_bit(NFS_LAYOUT_RETURN_REQUESTED,
Peng Taoe755d632015-08-19 13:49:19 +08001104 &lo->plh_flags))
1105 layoutreturn = pnfs_prepare_layoutreturn(lo);
1106
Fred Isamanf7e89172011-01-06 11:36:32 +00001107 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
Peng Taoe755d632015-08-19 13:49:19 +08001108 /* If we are sending layoutreturn, invalidate all valid lsegs */
1109 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
Fred Isamanf7e89172011-01-06 11:36:32 +00001110 mark_lseg_invalid(lseg, &tmp_list);
1111 found = true;
1112 }
Peng Tao500d7012015-09-22 11:35:22 +08001113 /* ROC in two conditions:
Peng Taoe755d632015-08-19 13:49:19 +08001114 * 1. there are ROC lsegs
1115 * 2. we don't send layoutreturn
Peng Taoe755d632015-08-19 13:49:19 +08001116 */
Peng Tao500d7012015-09-22 11:35:22 +08001117 if (found && !layoutreturn) {
1118 /* lo ref dropped in pnfs_roc_release() */
1119 pnfs_get_layout_hdr(lo);
Peng Taoe755d632015-08-19 13:49:19 +08001120 roc = true;
Peng Tao500d7012015-09-22 11:35:22 +08001121 }
Peng Taoe755d632015-08-19 13:49:19 +08001122
1123out_noroc:
Fred Isamanf7e89172011-01-06 11:36:32 +00001124 spin_unlock(&ino->i_lock);
1125 pnfs_free_lseg_list(&tmp_list);
Trond Myklebust71401712015-03-25 12:36:13 -04001126 pnfs_layoutcommit_inode(ino, true);
Peng Taoe755d632015-08-19 13:49:19 +08001127 if (layoutreturn)
Trond Myklebusted429d62016-01-04 12:30:55 -05001128 pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
Peng Taoe755d632015-08-19 13:49:19 +08001129 return roc;
Fred Isamanf7e89172011-01-06 11:36:32 +00001130}
1131
1132void pnfs_roc_release(struct inode *ino)
1133{
1134 struct pnfs_layout_hdr *lo;
1135
1136 spin_lock(&ino->i_lock);
1137 lo = NFS_I(ino)->layout;
Trond Myklebust5c4a79f2015-08-04 15:57:13 -04001138 pnfs_clear_layoutreturn_waitbit(lo);
Trond Myklebust6622c3e2012-09-20 17:23:11 -04001139 if (atomic_dec_and_test(&lo->plh_refcount)) {
1140 pnfs_detach_layout_hdr(lo);
1141 spin_unlock(&ino->i_lock);
1142 pnfs_free_layout_hdr(lo);
1143 } else
1144 spin_unlock(&ino->i_lock);
Fred Isamanf7e89172011-01-06 11:36:32 +00001145}
1146
1147void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1148{
1149 struct pnfs_layout_hdr *lo;
1150
1151 spin_lock(&ino->i_lock);
1152 lo = NFS_I(ino)->layout;
Trond Myklebust0654cc72015-12-28 11:48:14 -05001153 pnfs_mark_layout_returned_if_empty(lo);
Trond Myklebust0f35ad62012-10-04 16:28:17 -07001154 if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
Fred Isamanf7e89172011-01-06 11:36:32 +00001155 lo->plh_barrier = barrier;
1156 spin_unlock(&ino->i_lock);
Trond Myklebust6a463beb2015-08-20 15:40:47 -05001157 trace_nfs4_layoutreturn_on_close(ino, 0);
Fred Isamanf7e89172011-01-06 11:36:32 +00001158}
1159
Trond Myklebust4ff376f2015-08-18 23:23:21 -05001160void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
Fred Isamanf7e89172011-01-06 11:36:32 +00001161{
1162 struct nfs_inode *nfsi = NFS_I(ino);
Trond Myklebust7fdab062012-09-20 20:15:57 -04001163 struct pnfs_layout_hdr *lo;
Trond Myklebust7fdab062012-09-20 20:15:57 -04001164 u32 current_seqid;
Fred Isamanf7e89172011-01-06 11:36:32 +00001165
1166 spin_lock(&ino->i_lock);
Trond Myklebust7fdab062012-09-20 20:15:57 -04001167 lo = nfsi->layout;
1168 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
Fred Isamanf7e89172011-01-06 11:36:32 +00001169
Trond Myklebust7fdab062012-09-20 20:15:57 -04001170 /* Since close does not return a layout stateid for use as
1171 * a barrier, we choose the worst-case barrier.
1172 */
1173 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
Fred Isamanf7e89172011-01-06 11:36:32 +00001174 spin_unlock(&ino->i_lock);
Fred Isamanf7e89172011-01-06 11:36:32 +00001175}
1176
Peng Tao500d7012015-09-22 11:35:22 +08001177bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1178{
1179 struct nfs_inode *nfsi = NFS_I(ino);
1180 struct pnfs_layout_hdr *lo;
1181 bool sleep = false;
1182
1183 /* we might not have grabbed lo reference. so need to check under
1184 * i_lock */
1185 spin_lock(&ino->i_lock);
1186 lo = nfsi->layout;
1187 if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
1188 sleep = true;
1189 spin_unlock(&ino->i_lock);
1190
1191 if (sleep)
1192 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1193
1194 return sleep;
1195}
1196
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001197/*
1198 * Compare two layout segments for sorting into layout cache.
1199 * We want to preferentially return RW over RO layouts, so ensure those
1200 * are seen first.
1201 */
1202static s64
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001203pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001204 const struct pnfs_layout_range *l2)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001205{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001206 s64 d;
1207
1208 /* high offset > low offset */
1209 d = l1->offset - l2->offset;
1210 if (d)
1211 return d;
1212
1213 /* short length > long length */
1214 d = l2->length - l1->length;
1215 if (d)
1216 return d;
1217
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001218 /* read > read/write */
Benny Halevyfb3296e2011-05-22 19:47:26 +03001219 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001220}
1221
Trond Myklebust03772d22015-08-25 08:54:17 -04001222static bool
1223pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1224 const struct pnfs_layout_range *l2)
Andy Adamson974cec82010-10-20 00:18:02 -04001225{
Trond Myklebust03772d22015-08-25 08:54:17 -04001226 return pnfs_lseg_range_cmp(l1, l2) > 0;
1227}
1228
1229static bool
1230pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1231 struct pnfs_layout_segment *old)
1232{
1233 return false;
1234}
1235
1236void
1237pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1238 struct pnfs_layout_segment *lseg,
1239 bool (*is_after)(const struct pnfs_layout_range *,
1240 const struct pnfs_layout_range *),
1241 bool (*do_merge)(struct pnfs_layout_segment *,
1242 struct pnfs_layout_segment *),
1243 struct list_head *free_me)
1244{
1245 struct pnfs_layout_segment *lp, *tmp;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001246
Andy Adamson974cec82010-10-20 00:18:02 -04001247 dprintk("%s:Begin\n", __func__);
1248
Trond Myklebust03772d22015-08-25 08:54:17 -04001249 list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1250 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1251 continue;
1252 if (do_merge(lseg, lp)) {
1253 mark_lseg_invalid(lp, free_me);
1254 continue;
1255 }
1256 if (is_after(&lseg->pls_range, &lp->pls_range))
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001257 continue;
Fred Isaman566052c2011-01-06 11:36:20 +00001258 list_add_tail(&lseg->pls_list, &lp->pls_list);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001259 dprintk("%s: inserted lseg %p "
1260 "iomode %d offset %llu length %llu before "
1261 "lp %p iomode %d offset %llu length %llu\n",
Fred Isaman566052c2011-01-06 11:36:20 +00001262 __func__, lseg, lseg->pls_range.iomode,
1263 lseg->pls_range.offset, lseg->pls_range.length,
1264 lp, lp->pls_range.iomode, lp->pls_range.offset,
1265 lp->pls_range.length);
Benny Halevyfb3296e2011-05-22 19:47:26 +03001266 goto out;
Andy Adamson974cec82010-10-20 00:18:02 -04001267 }
Benny Halevyfb3296e2011-05-22 19:47:26 +03001268 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1269 dprintk("%s: inserted lseg %p "
1270 "iomode %d offset %llu length %llu at tail\n",
1271 __func__, lseg, lseg->pls_range.iomode,
1272 lseg->pls_range.offset, lseg->pls_range.length);
1273out:
Trond Myklebust70c3bd22012-09-18 20:51:13 -04001274 pnfs_get_layout_hdr(lo);
Andy Adamson974cec82010-10-20 00:18:02 -04001275
1276 dprintk("%s:Return\n", __func__);
Benny Halevye5e94012010-10-20 00:18:01 -04001277}
Trond Myklebust03772d22015-08-25 08:54:17 -04001278EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1279
1280static void
1281pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1282 struct pnfs_layout_segment *lseg,
1283 struct list_head *free_me)
1284{
1285 struct inode *inode = lo->plh_inode;
1286 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1287
1288 if (ld->add_lseg != NULL)
1289 ld->add_lseg(lo, lseg, free_me);
1290 else
1291 pnfs_generic_layout_insert_lseg(lo, lseg,
1292 pnfs_lseg_range_is_after,
1293 pnfs_lseg_no_merge,
1294 free_me);
1295}
Benny Halevye5e94012010-10-20 00:18:01 -04001296
1297static struct pnfs_layout_hdr *
Peng Tao9fa40752011-07-30 20:52:32 -04001298alloc_init_layout_hdr(struct inode *ino,
1299 struct nfs_open_context *ctx,
1300 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001301{
1302 struct pnfs_layout_hdr *lo;
1303
Benny Halevy636fb9c2011-05-22 19:51:33 +03001304 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
Benny Halevye5e94012010-10-20 00:18:01 -04001305 if (!lo)
1306 return NULL;
Fred Isamancc6e5342011-01-06 11:36:28 +00001307 atomic_set(&lo->plh_refcount, 1);
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001308 INIT_LIST_HEAD(&lo->plh_layouts);
1309 INIT_LIST_HEAD(&lo->plh_segs);
Trond Myklebustfd9a8d72013-02-12 09:48:42 -05001310 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001311 lo->plh_inode = ino;
Trond Myklebust5cc22162013-05-21 09:26:49 -04001312 lo->plh_lc_cred = get_rpccred(ctx->cred);
Benny Halevye5e94012010-10-20 00:18:01 -04001313 return lo;
1314}
1315
1316static struct pnfs_layout_hdr *
Peng Tao9fa40752011-07-30 20:52:32 -04001317pnfs_find_alloc_layout(struct inode *ino,
1318 struct nfs_open_context *ctx,
1319 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001320{
1321 struct nfs_inode *nfsi = NFS_I(ino);
1322 struct pnfs_layout_hdr *new = NULL;
1323
1324 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1325
Trond Myklebust251ec412012-10-02 15:41:05 -07001326 if (nfsi->layout != NULL)
1327 goto out_existing;
Benny Halevye5e94012010-10-20 00:18:01 -04001328 spin_unlock(&ino->i_lock);
Peng Tao9fa40752011-07-30 20:52:32 -04001329 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
Benny Halevye5e94012010-10-20 00:18:01 -04001330 spin_lock(&ino->i_lock);
1331
Trond Myklebust251ec412012-10-02 15:41:05 -07001332 if (likely(nfsi->layout == NULL)) { /* Won the race? */
Benny Halevye5e94012010-10-20 00:18:01 -04001333 nfsi->layout = new;
Trond Myklebust251ec412012-10-02 15:41:05 -07001334 return new;
Yanchuan Nian7175fe92012-10-31 16:05:48 +08001335 } else if (new != NULL)
1336 pnfs_free_layout_hdr(new);
Trond Myklebust251ec412012-10-02 15:41:05 -07001337out_existing:
1338 pnfs_get_layout_hdr(nfsi->layout);
Benny Halevye5e94012010-10-20 00:18:01 -04001339 return nfsi->layout;
1340}
1341
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001342/*
1343 * iomode matching rules:
1344 * iomode lseg match
1345 * ----- ----- -----
1346 * ANY READ true
1347 * ANY RW true
1348 * RW READ false
1349 * RW RW true
1350 * READ READ true
1351 * READ RW true
1352 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001353static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001354pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001355 const struct pnfs_layout_range *range)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001356{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001357 struct pnfs_layout_range range1;
1358
1359 if ((range->iomode == IOMODE_RW &&
1360 ls_range->iomode != IOMODE_RW) ||
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001361 !pnfs_lseg_range_intersecting(ls_range, range))
Benny Halevyfb3296e2011-05-22 19:47:26 +03001362 return 0;
1363
1364 /* range1 covers only the first byte in the range */
1365 range1 = *range;
1366 range1.length = 1;
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001367 return pnfs_lseg_range_contained(ls_range, &range1);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001368}
1369
1370/*
1371 * lookup range in layout
1372 */
Benny Halevye5e94012010-10-20 00:18:01 -04001373static struct pnfs_layout_segment *
Benny Halevyfb3296e2011-05-22 19:47:26 +03001374pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1375 struct pnfs_layout_range *range)
Benny Halevye5e94012010-10-20 00:18:01 -04001376{
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001377 struct pnfs_layout_segment *lseg, *ret = NULL;
1378
1379 dprintk("%s:Begin\n", __func__);
1380
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001381 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
Fred Isaman4541d162011-01-06 11:36:23 +00001382 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
Peng Taoce6ab4f2014-09-06 00:53:24 +08001383 !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001384 pnfs_lseg_range_match(&lseg->pls_range, range)) {
Trond Myklebust9369a432012-09-18 20:57:08 -04001385 ret = pnfs_get_lseg(lseg);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001386 break;
1387 }
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001388 }
1389
1390 dprintk("%s:Return lseg %p ref %d\n",
Fred Isaman4541d162011-01-06 11:36:23 +00001391 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001392 return ret;
Benny Halevye5e94012010-10-20 00:18:01 -04001393}
1394
1395/*
Andy Adamsond23d61c2012-05-23 05:02:37 -04001396 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1397 * to the MDS or over pNFS
1398 *
1399 * The nfs_inode read_io and write_io fields are cumulative counters reset
1400 * when there are no layout segments. Note that in pnfs_update_layout iomode
1401 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1402 * WRITE request.
1403 *
1404 * A return of true means use MDS I/O.
1405 *
1406 * From rfc 5661:
1407 * If a file's size is smaller than the file size threshold, data accesses
1408 * SHOULD be sent to the metadata server. If an I/O request has a length that
1409 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1410 * server. If both file size and I/O size are provided, the client SHOULD
1411 * reach or exceed both thresholds before sending its read or write
1412 * requests to the data server.
1413 */
1414static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1415 struct inode *ino, int iomode)
1416{
1417 struct nfs4_threshold *t = ctx->mdsthreshold;
1418 struct nfs_inode *nfsi = NFS_I(ino);
1419 loff_t fsize = i_size_read(ino);
1420 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1421
1422 if (t == NULL)
1423 return ret;
1424
1425 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1426 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1427
1428 switch (iomode) {
1429 case IOMODE_READ:
1430 if (t->bm & THRESHOLD_RD) {
1431 dprintk("%s fsize %llu\n", __func__, fsize);
1432 size_set = true;
1433 if (fsize < t->rd_sz)
1434 size = true;
1435 }
1436 if (t->bm & THRESHOLD_RD_IO) {
1437 dprintk("%s nfsi->read_io %llu\n", __func__,
1438 nfsi->read_io);
1439 io_set = true;
1440 if (nfsi->read_io < t->rd_io_sz)
1441 io = true;
1442 }
1443 break;
1444 case IOMODE_RW:
1445 if (t->bm & THRESHOLD_WR) {
1446 dprintk("%s fsize %llu\n", __func__, fsize);
1447 size_set = true;
1448 if (fsize < t->wr_sz)
1449 size = true;
1450 }
1451 if (t->bm & THRESHOLD_WR_IO) {
1452 dprintk("%s nfsi->write_io %llu\n", __func__,
1453 nfsi->write_io);
1454 io_set = true;
1455 if (nfsi->write_io < t->wr_io_sz)
1456 io = true;
1457 }
1458 break;
1459 }
1460 if (size_set && io_set) {
1461 if (size && io)
1462 ret = true;
1463 } else if (size || io)
1464 ret = true;
1465
1466 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1467 return ret;
1468}
1469
Peng Taoaa8a45e2014-12-01 08:22:23 +08001470static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1471{
1472 /*
1473 * send layoutcommit as it can hold up layoutreturn due to lseg
1474 * reference
1475 */
1476 pnfs_layoutcommit_inode(lo->plh_inode, false);
1477 return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
Trond Myklebust2e5b29f2015-12-14 16:25:11 -05001478 nfs_wait_bit_killable,
Peng Taoaa8a45e2014-12-01 08:22:23 +08001479 TASK_UNINTERRUPTIBLE);
1480}
1481
Tom Haynesd67ae822014-12-11 17:02:04 -05001482static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1483{
1484 unsigned long *bitlock = &lo->plh_flags;
1485
1486 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1487 smp_mb__after_atomic();
1488 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1489}
1490
Andy Adamsond23d61c2012-05-23 05:02:37 -04001491/*
Benny Halevye5e94012010-10-20 00:18:01 -04001492 * Layout segment is retreived from the server if not cached.
1493 * The appropriate layout segment is referenced and returned to the caller.
1494 */
Andy Adamson7c24d942011-06-13 18:22:38 -04001495struct pnfs_layout_segment *
Benny Halevye5e94012010-10-20 00:18:01 -04001496pnfs_update_layout(struct inode *ino,
1497 struct nfs_open_context *ctx,
Benny Halevyfb3296e2011-05-22 19:47:26 +03001498 loff_t pos,
1499 u64 count,
Trond Myklebusta75b9df2011-05-11 18:00:51 -04001500 enum pnfs_iomode iomode,
1501 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001502{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001503 struct pnfs_layout_range arg = {
1504 .iomode = iomode,
1505 .offset = pos,
1506 .length = count,
1507 };
Benny Halevy707ed5f2011-05-22 19:47:46 +03001508 unsigned pg_offset;
Weston Andros Adamson6382a442011-06-01 16:44:44 -04001509 struct nfs_server *server = NFS_SERVER(ino);
1510 struct nfs_client *clp = server->nfs_client;
Benny Halevye5e94012010-10-20 00:18:01 -04001511 struct pnfs_layout_hdr *lo;
1512 struct pnfs_layout_segment *lseg = NULL;
Weston Andros Adamson30005122013-02-28 20:30:10 -05001513 bool first;
Benny Halevye5e94012010-10-20 00:18:01 -04001514
Jeff Layton9a4bf312015-12-10 10:41:58 -05001515 if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001516 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001517 PNFS_UPDATE_LAYOUT_NO_PNFS);
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001518 goto out;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001519 }
Andy Adamsond23d61c2012-05-23 05:02:37 -04001520
Jeff Layton9a4bf312015-12-10 10:41:58 -05001521 if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001522 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001523 PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
Trond Myklebust4ae935602015-08-31 01:25:11 -07001524 goto out;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001525 }
Trond Myklebust4ae935602015-08-31 01:25:11 -07001526
Jeff Layton9a4bf312015-12-10 10:41:58 -05001527 if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001528 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001529 PNFS_UPDATE_LAYOUT_MDSTHRESH);
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001530 goto out;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001531 }
Andy Adamsond23d61c2012-05-23 05:02:37 -04001532
Peng Tao9bf87482014-08-22 17:37:41 +08001533lookup_again:
1534 first = false;
Benny Halevye5e94012010-10-20 00:18:01 -04001535 spin_lock(&ino->i_lock);
Peng Tao9fa40752011-07-30 20:52:32 -04001536 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001537 if (lo == NULL) {
1538 spin_unlock(&ino->i_lock);
Trond Myklebustf4848302015-12-26 18:03:07 -05001539 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001540 PNFS_UPDATE_LAYOUT_NOMEM);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001541 goto out;
1542 }
Benny Halevye5e94012010-10-20 00:18:01 -04001543
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001544 /* Do we even need to bother with this? */
Trond Myklebusta59c30a2012-03-01 11:17:47 -05001545 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001546 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001547 PNFS_UPDATE_LAYOUT_BULK_RECALL);
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001548 dprintk("%s matches recall, use MDS\n", __func__);
Benny Halevye5e94012010-10-20 00:18:01 -04001549 goto out_unlock;
1550 }
1551
1552 /* if LAYOUTGET already failed once we don't try again */
Trond Myklebust2e5b29f2015-12-14 16:25:11 -05001553 if (pnfs_layout_io_test_failed(lo, iomode)) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001554 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001555 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
Benny Halevye5e94012010-10-20 00:18:01 -04001556 goto out_unlock;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001557 }
Benny Halevye5e94012010-10-20 00:18:01 -04001558
Peng Tao9bf87482014-08-22 17:37:41 +08001559 first = list_empty(&lo->plh_segs);
1560 if (first) {
1561 /* The first layoutget for the file. Need to serialize per
1562 * RFC 5661 Errata 3208.
1563 */
1564 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1565 &lo->plh_flags)) {
1566 spin_unlock(&ino->i_lock);
1567 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1568 TASK_UNINTERRUPTIBLE);
1569 pnfs_put_layout_hdr(lo);
1570 goto lookup_again;
1571 }
1572 } else {
1573 /* Check to see if the layout for the given range
1574 * already exists
1575 */
1576 lseg = pnfs_find_lseg(lo, &arg);
Jeff Layton9a4bf312015-12-10 10:41:58 -05001577 if (lseg) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001578 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001579 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
Peng Tao9bf87482014-08-22 17:37:41 +08001580 goto out_unlock;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001581 }
Peng Tao9bf87482014-08-22 17:37:41 +08001582 }
Andy Adamson568e8c42011-03-01 01:34:22 +00001583
Peng Taoaa8a45e2014-12-01 08:22:23 +08001584 /*
1585 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1586 * for LAYOUTRETURN even if first is true.
1587 */
Trond Myklebust8f70f532015-08-04 16:09:44 -04001588 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
Peng Taoaa8a45e2014-12-01 08:22:23 +08001589 spin_unlock(&ino->i_lock);
1590 dprintk("%s wait for layoutreturn\n", __func__);
1591 if (pnfs_prepare_to_retry_layoutget(lo)) {
Tom Haynesd67ae822014-12-11 17:02:04 -05001592 if (first)
1593 pnfs_clear_first_layoutget(lo);
Peng Taoaa8a45e2014-12-01 08:22:23 +08001594 pnfs_put_layout_hdr(lo);
1595 dprintk("%s retrying\n", __func__);
1596 goto lookup_again;
1597 }
Trond Myklebustf4848302015-12-26 18:03:07 -05001598 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001599 PNFS_UPDATE_LAYOUT_RETURN);
Peng Taoaa8a45e2014-12-01 08:22:23 +08001600 goto out_put_layout_hdr;
1601 }
1602
Jeff Layton9a4bf312015-12-10 10:41:58 -05001603 if (pnfs_layoutgets_blocked(lo)) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001604 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001605 PNFS_UPDATE_LAYOUT_BLOCKED);
Fred Isamancf7d63f2011-01-06 11:36:25 +00001606 goto out_unlock;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001607 }
Fred Isamancf7d63f2011-01-06 11:36:25 +00001608 atomic_inc(&lo->plh_outstanding);
Fred Isamanf49f9ba2011-02-03 18:28:52 +00001609 spin_unlock(&ino->i_lock);
Weston Andros Adamson30005122013-02-28 20:30:10 -05001610
Peng Taoabb9a002014-08-22 17:37:40 +08001611 if (list_empty(&lo->plh_layouts)) {
Fred Isaman2130ff62011-01-06 11:36:26 +00001612 /* The lo must be on the clp list if there is any
1613 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1614 */
1615 spin_lock(&clp->cl_lock);
Peng Taoabb9a002014-08-22 17:37:40 +08001616 if (list_empty(&lo->plh_layouts))
1617 list_add_tail(&lo->plh_layouts, &server->layouts);
Fred Isaman2130ff62011-01-06 11:36:26 +00001618 spin_unlock(&clp->cl_lock);
1619 }
Benny Halevye5e94012010-10-20 00:18:01 -04001620
Benny Halevy707ed5f2011-05-22 19:47:46 +03001621 pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1622 if (pg_offset) {
1623 arg.offset -= pg_offset;
1624 arg.length += pg_offset;
1625 }
Andy Adamson7c24d942011-06-13 18:22:38 -04001626 if (arg.length != NFS4_MAX_UINT64)
1627 arg.length = PAGE_CACHE_ALIGN(arg.length);
Benny Halevy707ed5f2011-05-22 19:47:46 +03001628
Benny Halevyfb3296e2011-05-22 19:47:26 +03001629 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
Fred Isamancf7d63f2011-01-06 11:36:25 +00001630 atomic_dec(&lo->plh_outstanding);
Trond Myklebustf4848302015-12-26 18:03:07 -05001631 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001632 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001633out_put_layout_hdr:
Tom Haynesd67ae822014-12-11 17:02:04 -05001634 if (first)
1635 pnfs_clear_first_layoutget(lo);
Trond Myklebust70c3bd22012-09-18 20:51:13 -04001636 pnfs_put_layout_hdr(lo);
Benny Halevye5e94012010-10-20 00:18:01 -04001637out:
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001638 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1639 "(%s, offset: %llu, length: %llu)\n",
1640 __func__, ino->i_sb->s_id,
1641 (unsigned long long)NFS_FILEID(ino),
Peng Taod600ad12015-12-04 02:57:48 +08001642 IS_ERR_OR_NULL(lseg) ? "not found" : "found",
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001643 iomode==IOMODE_RW ? "read/write" : "read-only",
1644 (unsigned long long)pos,
1645 (unsigned long long)count);
Benny Halevye5e94012010-10-20 00:18:01 -04001646 return lseg;
1647out_unlock:
1648 spin_unlock(&ino->i_lock);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001649 goto out_put_layout_hdr;
Benny Halevye5e94012010-10-20 00:18:01 -04001650}
Andy Adamson7c24d942011-06-13 18:22:38 -04001651EXPORT_SYMBOL_GPL(pnfs_update_layout);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001652
Trond Myklebust540d9862015-08-25 11:16:13 -04001653static bool
1654pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1655{
1656 switch (range->iomode) {
1657 case IOMODE_READ:
1658 case IOMODE_RW:
1659 break;
1660 default:
1661 return false;
1662 }
1663 if (range->offset == NFS4_MAX_UINT64)
1664 return false;
1665 if (range->length == 0)
1666 return false;
1667 if (range->length != NFS4_MAX_UINT64 &&
1668 range->length > NFS4_MAX_UINT64 - range->offset)
1669 return false;
1670 return true;
1671}
1672
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001673struct pnfs_layout_segment *
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001674pnfs_layout_process(struct nfs4_layoutget *lgp)
1675{
1676 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1677 struct nfs4_layoutget_res *res = &lgp->res;
1678 struct pnfs_layout_segment *lseg;
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001679 struct inode *ino = lo->plh_inode;
Trond Myklebust78096cc2014-02-12 10:02:27 -05001680 LIST_HEAD(free_me);
Trond Myklebust540d9862015-08-25 11:16:13 -04001681 int status = -EINVAL;
1682
1683 if (!pnfs_sanity_check_layout_range(&res->range))
1684 goto out;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001685
1686 /* Inject layout blob into I/O device driver */
Trond Myklebusta75b9df2011-05-11 18:00:51 -04001687 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001688 if (!lseg || IS_ERR(lseg)) {
1689 if (!lseg)
1690 status = -ENOMEM;
1691 else
1692 status = PTR_ERR(lseg);
1693 dprintk("%s: Could not allocate layout: error %d\n",
1694 __func__, status);
1695 goto out;
1696 }
1697
Christoph Hellwig1013df612014-08-21 11:09:18 -05001698 init_lseg(lo, lseg);
1699 lseg->pls_range = res->range;
1700
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001701 spin_lock(&ino->i_lock);
Trond Myklebuste1c06f82015-08-04 16:40:08 -04001702 if (pnfs_layoutgets_blocked(lo)) {
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001703 dprintk("%s forget reply due to state\n", __func__);
1704 goto out_forget_reply;
1705 }
Trond Myklebust038d6492012-10-02 16:38:41 -07001706
Christoph Hellwig362f7472014-08-21 11:09:20 -05001707 if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1708 /* existing state ID, make sure the sequence number matches. */
1709 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1710 dprintk("%s forget reply due to sequence\n", __func__);
Jeff Layton4f2e9dc2015-11-25 13:43:14 -05001711 status = -EAGAIN;
Christoph Hellwig362f7472014-08-21 11:09:20 -05001712 goto out_forget_reply;
1713 }
1714 pnfs_set_layout_stateid(lo, &res->stateid, false);
1715 } else {
1716 /*
1717 * We got an entirely new state ID. Mark all segments for the
1718 * inode invalid, and don't bother validating the stateid
1719 * sequence number.
1720 */
1721 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1722
1723 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1724 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1725 }
Trond Myklebust038d6492012-10-02 16:38:41 -07001726
Christoph Hellwig47abade2014-08-21 11:09:22 -05001727 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1728
Trond Myklebust9369a432012-09-18 20:57:08 -04001729 pnfs_get_lseg(lseg);
Trond Myklebust03772d22015-08-25 08:54:17 -04001730 pnfs_layout_insert_lseg(lo, lseg, &free_me);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001731
Peng Tao39761432015-08-21 12:49:44 +08001732 if (res->return_on_close)
Fred Isamanf7e89172011-01-06 11:36:32 +00001733 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
Fred Isamanf7e89172011-01-06 11:36:32 +00001734
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001735 spin_unlock(&ino->i_lock);
Trond Myklebust78096cc2014-02-12 10:02:27 -05001736 pnfs_free_lseg_list(&free_me);
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001737 return lseg;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001738out:
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001739 return ERR_PTR(status);
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001740
1741out_forget_reply:
1742 spin_unlock(&ino->i_lock);
1743 lseg->pls_layout = lo;
1744 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1745 goto out;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001746}
1747
Peng Tao016256d2014-09-06 00:53:23 +08001748static void
Trond Myklebust5c97f5d2016-01-04 11:53:04 -05001749pnfs_set_plh_return_iomode(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode)
1750{
1751 if (lo->plh_return_iomode == iomode)
1752 return;
1753 if (lo->plh_return_iomode != 0)
1754 iomode = IOMODE_ANY;
1755 lo->plh_return_iomode = iomode;
Trond Myklebuste0fa0d02016-02-15 12:56:17 -05001756 set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
Trond Myklebust5c97f5d2016-01-04 11:53:04 -05001757}
1758
Trond Myklebust2f215962016-02-15 12:36:04 -05001759/**
1760 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1761 * @lo: pointer to layout header
1762 * @tmp_list: list header to be used with pnfs_free_lseg_list()
1763 * @return_range: describe layout segment ranges to be returned
1764 *
1765 * This function is mainly intended for use by layoutrecall. It attempts
1766 * to free the layout segment immediately, or else to mark it for return
1767 * as soon as its reference count drops to zero.
1768 */
Trond Myklebust10335552016-01-04 11:23:52 -05001769int
Peng Tao016256d2014-09-06 00:53:23 +08001770pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1771 struct list_head *tmp_list,
Trond Myklebust506c0d62016-01-04 13:04:47 -05001772 const struct pnfs_layout_range *return_range)
Peng Tao016256d2014-09-06 00:53:23 +08001773{
1774 struct pnfs_layout_segment *lseg, *next;
Trond Myklebust10335552016-01-04 11:23:52 -05001775 int remaining = 0;
Peng Tao016256d2014-09-06 00:53:23 +08001776
1777 dprintk("%s:Begin lo %p\n", __func__, lo);
1778
1779 if (list_empty(&lo->plh_segs))
Trond Myklebust10335552016-01-04 11:23:52 -05001780 return 0;
Peng Tao016256d2014-09-06 00:53:23 +08001781
Trond Myklebustfc7ff362015-12-28 10:28:59 -05001782 assert_spin_locked(&lo->plh_inode->i_lock);
Peng Tao016256d2014-09-06 00:53:23 +08001783
1784 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1785 if (should_free_lseg(&lseg->pls_range, return_range)) {
1786 dprintk("%s: marking lseg %p iomode %d "
1787 "offset %llu length %llu\n", __func__,
1788 lseg, lseg->pls_range.iomode,
1789 lseg->pls_range.offset,
1790 lseg->pls_range.length);
Trond Myklebust2f215962016-02-15 12:36:04 -05001791 if (mark_lseg_invalid(lseg, tmp_list))
1792 continue;
1793 remaining++;
Peng Tao016256d2014-09-06 00:53:23 +08001794 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
Trond Myklebust5c97f5d2016-01-04 11:53:04 -05001795 pnfs_set_plh_return_iomode(lo, return_range->iomode);
Peng Tao016256d2014-09-06 00:53:23 +08001796 }
Trond Myklebust10335552016-01-04 11:23:52 -05001797 return remaining;
Peng Tao016256d2014-09-06 00:53:23 +08001798}
1799
1800void pnfs_error_mark_layout_for_return(struct inode *inode,
1801 struct pnfs_layout_segment *lseg)
1802{
1803 struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
Peng Tao016256d2014-09-06 00:53:23 +08001804 struct pnfs_layout_range range = {
1805 .iomode = lseg->pls_range.iomode,
1806 .offset = 0,
1807 .length = NFS4_MAX_UINT64,
1808 };
1809 LIST_HEAD(free_me);
Trond Myklebust10335552016-01-04 11:23:52 -05001810 bool return_now = false;
Peng Tao016256d2014-09-06 00:53:23 +08001811
1812 spin_lock(&inode->i_lock);
Trond Myklebust5c97f5d2016-01-04 11:53:04 -05001813 pnfs_set_plh_return_iomode(lo, range.iomode);
Peng Tao016256d2014-09-06 00:53:23 +08001814 /*
1815 * mark all matching lsegs so that we are sure to have no live
1816 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1817 * for how it works.
1818 */
Trond Myklebust10335552016-01-04 11:23:52 -05001819 if (!pnfs_mark_matching_lsegs_return(lo, &free_me, &range)) {
1820 nfs4_stateid stateid;
1821 enum pnfs_iomode iomode = lo->plh_return_iomode;
1822
1823 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1824 return_now = pnfs_prepare_layoutreturn(lo);
1825 spin_unlock(&inode->i_lock);
1826 if (return_now)
1827 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1828 } else {
1829 spin_unlock(&inode->i_lock);
1830 nfs_commit_inode(inode, 0);
1831 }
Peng Tao016256d2014-09-06 00:53:23 +08001832 pnfs_free_lseg_list(&free_me);
1833}
1834EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1835
Trond Myklebustd8007d42011-06-10 13:30:23 -04001836void
1837pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1838{
Peng Tao1fd937b2012-09-25 14:55:57 +08001839 u64 rd_size = req->wb_bytes;
1840
Peng Taocb5d04b2015-01-24 22:14:52 +08001841 if (pgio->pg_lseg == NULL) {
1842 if (pgio->pg_dreq == NULL)
1843 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1844 else
1845 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
Trond Myklebustd8007d42011-06-10 13:30:23 -04001846
Peng Taocb5d04b2015-01-24 22:14:52 +08001847 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1848 req->wb_context,
1849 req_offset(req),
1850 rd_size,
1851 IOMODE_READ,
1852 GFP_KERNEL);
Peng Taod600ad12015-12-04 02:57:48 +08001853 if (IS_ERR(pgio->pg_lseg)) {
1854 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1855 pgio->pg_lseg = NULL;
1856 return;
1857 }
Peng Taocb5d04b2015-01-24 22:14:52 +08001858 }
Trond Myklebuste885de12011-06-10 13:30:23 -04001859 /* If no lseg, fall back to read through mds */
1860 if (pgio->pg_lseg == NULL)
Trond Myklebust1f945352011-07-13 15:59:57 -04001861 nfs_pageio_reset_read_mds(pgio);
Trond Myklebuste885de12011-06-10 13:30:23 -04001862
Trond Myklebustd8007d42011-06-10 13:30:23 -04001863}
1864EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1865
1866void
Peng Tao62965562012-09-25 14:55:57 +08001867pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1868 struct nfs_page *req, u64 wb_size)
Trond Myklebustd8007d42011-06-10 13:30:23 -04001869{
Peng Taod600ad12015-12-04 02:57:48 +08001870 if (pgio->pg_lseg == NULL) {
Peng Taocb5d04b2015-01-24 22:14:52 +08001871 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1872 req->wb_context,
1873 req_offset(req),
1874 wb_size,
1875 IOMODE_RW,
1876 GFP_NOFS);
Peng Taod600ad12015-12-04 02:57:48 +08001877 if (IS_ERR(pgio->pg_lseg)) {
1878 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1879 pgio->pg_lseg = NULL;
1880 return;
1881 }
1882 }
Trond Myklebuste885de12011-06-10 13:30:23 -04001883 /* If no lseg, fall back to write through mds */
1884 if (pgio->pg_lseg == NULL)
Trond Myklebust1f945352011-07-13 15:59:57 -04001885 nfs_pageio_reset_write_mds(pgio);
Trond Myklebustd8007d42011-06-10 13:30:23 -04001886}
1887EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1888
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04001889void
1890pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1891{
1892 if (desc->pg_lseg) {
1893 pnfs_put_lseg(desc->pg_lseg);
1894 desc->pg_lseg = NULL;
1895 }
1896}
1897EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1898
Weston Andros Adamsonb4fdac12014-05-15 11:56:43 -04001899/*
1900 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1901 * of bytes (maximum @req->wb_bytes) that can be coalesced.
1902 */
1903size_t
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001904pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1905 struct nfs_page *prev, struct nfs_page *req)
Fred Isamanbae724e2011-03-01 01:34:15 +00001906{
Weston Andros Adamson0f9c429e2014-05-15 11:56:51 -04001907 unsigned int size;
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001908 u64 seg_end, req_start, seg_left;
Weston Andros Adamson0f9c429e2014-05-15 11:56:51 -04001909
1910 size = nfs_generic_pg_test(pgio, prev, req);
Weston Andros Adamson0f9c429e2014-05-15 11:56:51 -04001911 if (!size)
1912 return 0;
Benny Halevy89a58e32011-05-25 20:54:40 +03001913
Trond Myklebust19982ba2011-06-10 13:30:23 -04001914 /*
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001915 * 'size' contains the number of bytes left in the current page (up
1916 * to the original size asked for in @req->wb_bytes).
1917 *
1918 * Calculate how many bytes are left in the layout segment
1919 * and if there are less bytes than 'size', return that instead.
Trond Myklebust19982ba2011-06-10 13:30:23 -04001920 *
1921 * Please also note that 'end_offset' is actually the offset of the
1922 * first byte that lies outside the pnfs_layout_range. FIXME?
1923 *
1924 */
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001925 if (pgio->pg_lseg) {
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001926 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1927 pgio->pg_lseg->pls_range.length);
1928 req_start = req_offset(req);
Weston Andros Adamson7c137892015-01-30 11:01:02 -05001929 WARN_ON_ONCE(req_start >= seg_end);
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001930 /* start of request is past the last byte of this segment */
Weston Andros Adamson7c137892015-01-30 11:01:02 -05001931 if (req_start >= seg_end) {
1932 /* reference the new lseg */
1933 if (pgio->pg_ops->pg_cleanup)
1934 pgio->pg_ops->pg_cleanup(pgio);
1935 if (pgio->pg_ops->pg_init)
1936 pgio->pg_ops->pg_init(pgio, req);
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001937 return 0;
Weston Andros Adamson7c137892015-01-30 11:01:02 -05001938 }
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001939
1940 /* adjust 'size' iff there are fewer bytes left in the
1941 * segment than what nfs_generic_pg_test returned */
1942 seg_left = seg_end - req_start;
1943 if (seg_left < size)
1944 size = (unsigned int)seg_left;
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001945 }
Weston Andros Adamson0f9c429e2014-05-15 11:56:51 -04001946
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001947 return size;
Fred Isamanbae724e2011-03-01 01:34:15 +00001948}
Benny Halevy89a58e32011-05-25 20:54:40 +03001949EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
Fred Isamanbae724e2011-03-01 01:34:15 +00001950
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001951int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001952{
1953 struct nfs_pageio_descriptor pgio;
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001954
1955 /* Resend all requests through the MDS */
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001956 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1957 hdr->completion_ops);
Trond Myklebustc7070112015-06-17 19:56:22 -04001958 set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001959 return nfs_pageio_resend(&pgio, hdr);
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001960}
Andy Adamsone7dd79af2012-04-27 17:53:46 -04001961EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001962
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001963static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001964{
Fred Isamancd841602012-04-20 14:47:44 -04001965
1966 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1967 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001968 PNFS_LAYOUTRET_ON_ERROR) {
Fred Isamancd841602012-04-20 14:47:44 -04001969 pnfs_return_layout(hdr->inode);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001970 }
Fred Isaman6c75dc02012-04-20 14:47:47 -04001971 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001972 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001973}
1974
Benny Halevyd20581a2011-05-22 19:52:03 +03001975/*
1976 * Called by non rpc-based layout drivers
1977 */
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001978void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
Fred Isaman94ad1c82011-03-01 01:34:14 +00001979{
Kinglong Meef8417b482015-10-16 17:23:29 +08001980 if (likely(!hdr->pnfs_error)) {
Trond Myklebust67af7612015-03-25 20:40:38 -04001981 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1982 hdr->mds_offset + hdr->res.count);
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001983 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
Kinglong Meef8417b482015-10-16 17:23:29 +08001984 }
1985 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1986 if (unlikely(hdr->pnfs_error))
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001987 pnfs_ld_handle_write_error(hdr);
1988 hdr->mds_ops->rpc_release(hdr);
Fred Isaman44b83792011-03-03 15:13:44 +00001989}
Benny Halevyd20581a2011-05-22 19:52:03 +03001990EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
Fred Isaman44b83792011-03-03 15:13:44 +00001991
Trond Myklebustdce81292011-07-13 15:59:19 -04001992static void
1993pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001994 struct nfs_pgio_header *hdr)
Trond Myklebustdce81292011-07-13 15:59:19 -04001995{
Peng Tao48d635f2014-11-10 08:35:35 +08001996 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001997
Fred Isaman6c75dc02012-04-20 14:47:47 -04001998 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001999 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
Fred Isaman6c75dc02012-04-20 14:47:47 -04002000 nfs_pageio_reset_write_mds(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04002001 mirror->pg_recoalesce = 1;
Fred Isaman6c75dc02012-04-20 14:47:47 -04002002 }
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002003 nfs_pgio_data_destroy(hdr);
Trond Myklebust1ca018d2015-06-17 19:41:51 -04002004 hdr->release(hdr);
Trond Myklebustdce81292011-07-13 15:59:19 -04002005}
2006
2007static enum pnfs_try_status
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002008pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
Trond Myklebustdce81292011-07-13 15:59:19 -04002009 const struct rpc_call_ops *call_ops,
2010 struct pnfs_layout_segment *lseg,
2011 int how)
Andy Adamson0382b742011-03-03 15:13:45 +00002012{
Fred Isamancd841602012-04-20 14:47:44 -04002013 struct inode *inode = hdr->inode;
Andy Adamson0382b742011-03-03 15:13:45 +00002014 enum pnfs_try_status trypnfs;
2015 struct nfs_server *nfss = NFS_SERVER(inode);
2016
Fred Isamancd841602012-04-20 14:47:44 -04002017 hdr->mds_ops = call_ops;
Andy Adamson0382b742011-03-03 15:13:45 +00002018
2019 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002020 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2021 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
Fred Isaman6c75dc02012-04-20 14:47:47 -04002022 if (trypnfs != PNFS_NOT_ATTEMPTED)
Andy Adamson0382b742011-03-03 15:13:45 +00002023 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
Andy Adamson0382b742011-03-03 15:13:45 +00002024 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2025 return trypnfs;
2026}
2027
Trond Myklebustdce81292011-07-13 15:59:19 -04002028static void
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002029pnfs_do_write(struct nfs_pageio_descriptor *desc,
2030 struct nfs_pgio_header *hdr, int how)
Trond Myklebustdce81292011-07-13 15:59:19 -04002031{
Trond Myklebustdce81292011-07-13 15:59:19 -04002032 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2033 struct pnfs_layout_segment *lseg = desc->pg_lseg;
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002034 enum pnfs_try_status trypnfs;
Trond Myklebustdce81292011-07-13 15:59:19 -04002035
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002036 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002037 if (trypnfs == PNFS_NOT_ATTEMPTED)
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002038 pnfs_write_through_mds(desc, hdr);
Trond Myklebustdce81292011-07-13 15:59:19 -04002039}
2040
Fred Isaman6c75dc02012-04-20 14:47:47 -04002041static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2042{
Trond Myklebust9369a432012-09-18 20:57:08 -04002043 pnfs_put_lseg(hdr->lseg);
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04002044 nfs_pgio_header_free(hdr);
Fred Isaman6c75dc02012-04-20 14:47:47 -04002045}
2046
Trond Myklebustdce81292011-07-13 15:59:19 -04002047int
2048pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2049{
Fred Isaman6c75dc02012-04-20 14:47:47 -04002050 struct nfs_pgio_header *hdr;
Trond Myklebustdce81292011-07-13 15:59:19 -04002051 int ret;
2052
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04002053 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2054 if (!hdr) {
Peng Tao2bff2282015-12-05 02:03:17 +08002055 desc->pg_error = -ENOMEM;
2056 return desc->pg_error;
Fred Isaman6c75dc02012-04-20 14:47:47 -04002057 }
Fred Isaman6c75dc02012-04-20 14:47:47 -04002058 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04002059
Trond Myklebust9369a432012-09-18 20:57:08 -04002060 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
Anna Schumakeref2c4882014-05-06 09:12:36 -04002061 ret = nfs_generic_pgio(desc, hdr);
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04002062 if (!ret)
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002063 pnfs_do_write(desc, hdr, desc->pg_ioflags);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04002064
Fred Isaman6c75dc02012-04-20 14:47:47 -04002065 return ret;
Trond Myklebustdce81292011-07-13 15:59:19 -04002066}
2067EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2068
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04002069int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
Trond Myklebust62e4a762011-11-10 14:30:37 -05002070{
2071 struct nfs_pageio_descriptor pgio;
2072
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002073 /* Resend all requests through the MDS */
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04002074 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2075 return nfs_pageio_resend(&pgio, hdr);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002076}
Andy Adamsone7dd79af2012-04-27 17:53:46 -04002077EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002078
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002079static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002080{
Fred Isamancd841602012-04-20 14:47:44 -04002081 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2082 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002083 PNFS_LAYOUTRET_ON_ERROR) {
Fred Isamancd841602012-04-20 14:47:44 -04002084 pnfs_return_layout(hdr->inode);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002085 }
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002086 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04002087 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
Trond Myklebust62e4a762011-11-10 14:30:37 -05002088}
2089
Andy Adamsonb1f69b72010-10-20 00:18:03 -04002090/*
Benny Halevyd20581a2011-05-22 19:52:03 +03002091 * Called by non rpc-based layout drivers
2092 */
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002093void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
Benny Halevyd20581a2011-05-22 19:52:03 +03002094{
Fred Isamancd841602012-04-20 14:47:44 -04002095 if (likely(!hdr->pnfs_error)) {
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002096 __nfs4_read_done_cb(hdr);
2097 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
Kinglong Meef8417b482015-10-16 17:23:29 +08002098 }
2099 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2100 if (unlikely(hdr->pnfs_error))
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002101 pnfs_ld_handle_read_error(hdr);
2102 hdr->mds_ops->rpc_release(hdr);
Benny Halevyd20581a2011-05-22 19:52:03 +03002103}
2104EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2105
Trond Myklebust493292d2011-07-13 15:58:28 -04002106static void
2107pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002108 struct nfs_pgio_header *hdr)
Trond Myklebust493292d2011-07-13 15:58:28 -04002109{
Peng Tao48d635f2014-11-10 08:35:35 +08002110 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04002111
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002112 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04002113 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002114 nfs_pageio_reset_read_mds(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04002115 mirror->pg_recoalesce = 1;
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002116 }
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002117 nfs_pgio_data_destroy(hdr);
Trond Myklebust1ca018d2015-06-17 19:41:51 -04002118 hdr->release(hdr);
Trond Myklebust493292d2011-07-13 15:58:28 -04002119}
2120
Benny Halevyd20581a2011-05-22 19:52:03 +03002121/*
Andy Adamson64419a92011-03-01 01:34:16 +00002122 * Call the appropriate parallel I/O subsystem read function.
2123 */
Trond Myklebust493292d2011-07-13 15:58:28 -04002124static enum pnfs_try_status
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002125pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
Trond Myklebust493292d2011-07-13 15:58:28 -04002126 const struct rpc_call_ops *call_ops,
2127 struct pnfs_layout_segment *lseg)
Andy Adamson64419a92011-03-01 01:34:16 +00002128{
Fred Isamancd841602012-04-20 14:47:44 -04002129 struct inode *inode = hdr->inode;
Andy Adamson64419a92011-03-01 01:34:16 +00002130 struct nfs_server *nfss = NFS_SERVER(inode);
2131 enum pnfs_try_status trypnfs;
2132
Fred Isamancd841602012-04-20 14:47:44 -04002133 hdr->mds_ops = call_ops;
Andy Adamson64419a92011-03-01 01:34:16 +00002134
2135 dprintk("%s: Reading ino:%lu %u@%llu\n",
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002136 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
Andy Adamson64419a92011-03-01 01:34:16 +00002137
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002138 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002139 if (trypnfs != PNFS_NOT_ATTEMPTED)
Andy Adamson64419a92011-03-01 01:34:16 +00002140 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
Andy Adamson64419a92011-03-01 01:34:16 +00002141 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2142 return trypnfs;
2143}
Andy Adamson863a3c62011-03-23 13:27:54 +00002144
Peng Taoceb11e12014-11-10 08:35:38 +08002145/* Resend all requests through pnfs. */
2146int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2147{
2148 struct nfs_pageio_descriptor pgio;
2149
2150 nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
2151 return nfs_pageio_resend(&pgio, hdr);
2152}
2153EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2154
Trond Myklebust493292d2011-07-13 15:58:28 -04002155static void
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002156pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
Trond Myklebust493292d2011-07-13 15:58:28 -04002157{
Trond Myklebust493292d2011-07-13 15:58:28 -04002158 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2159 struct pnfs_layout_segment *lseg = desc->pg_lseg;
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002160 enum pnfs_try_status trypnfs;
Peng Taoceb11e12014-11-10 08:35:38 +08002161 int err = 0;
Trond Myklebust493292d2011-07-13 15:58:28 -04002162
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002163 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
Peng Taoceb11e12014-11-10 08:35:38 +08002164 if (trypnfs == PNFS_TRY_AGAIN)
2165 err = pnfs_read_resend_pnfs(hdr);
2166 if (trypnfs == PNFS_NOT_ATTEMPTED || err)
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002167 pnfs_read_through_mds(desc, hdr);
Trond Myklebust493292d2011-07-13 15:58:28 -04002168}
2169
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002170static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2171{
Trond Myklebust9369a432012-09-18 20:57:08 -04002172 pnfs_put_lseg(hdr->lseg);
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04002173 nfs_pgio_header_free(hdr);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002174}
2175
Trond Myklebust493292d2011-07-13 15:58:28 -04002176int
2177pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2178{
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002179 struct nfs_pgio_header *hdr;
Trond Myklebust493292d2011-07-13 15:58:28 -04002180 int ret;
2181
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04002182 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2183 if (!hdr) {
Peng Tao2bff2282015-12-05 02:03:17 +08002184 desc->pg_error = -ENOMEM;
2185 return desc->pg_error;
Trond Myklebust493292d2011-07-13 15:58:28 -04002186 }
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002187 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
Trond Myklebust9369a432012-09-18 20:57:08 -04002188 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
Anna Schumakeref2c4882014-05-06 09:12:36 -04002189 ret = nfs_generic_pgio(desc, hdr);
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04002190 if (!ret)
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002191 pnfs_do_read(desc, hdr);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002192 return ret;
Trond Myklebust493292d2011-07-13 15:58:28 -04002193}
2194EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2195
Trond Myklebust71244d92014-01-13 13:34:36 -05002196static void pnfs_clear_layoutcommitting(struct inode *inode)
2197{
2198 unsigned long *bitlock = &NFS_I(inode)->flags;
2199
2200 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
Peter Zijlstra4e857c52014-03-17 18:06:10 +01002201 smp_mb__after_atomic();
Trond Myklebust71244d92014-01-13 13:34:36 -05002202 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2203}
2204
Andy Adamson863a3c62011-03-23 13:27:54 +00002205/*
Peng Taoa9bae562011-07-30 20:52:33 -04002206 * There can be multiple RW segments.
Andy Adamson863a3c62011-03-23 13:27:54 +00002207 */
Peng Taoa9bae562011-07-30 20:52:33 -04002208static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
Andy Adamson863a3c62011-03-23 13:27:54 +00002209{
Peng Taoa9bae562011-07-30 20:52:33 -04002210 struct pnfs_layout_segment *lseg;
Andy Adamson863a3c62011-03-23 13:27:54 +00002211
Peng Taoa9bae562011-07-30 20:52:33 -04002212 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2213 if (lseg->pls_range.iomode == IOMODE_RW &&
Trond Myklebusta073dbf2013-03-20 12:34:32 -04002214 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
Peng Taoa9bae562011-07-30 20:52:33 -04002215 list_add(&lseg->pls_lc_list, listp);
2216 }
Andy Adamson863a3c62011-03-23 13:27:54 +00002217}
2218
Trond Myklebusta073dbf2013-03-20 12:34:32 -04002219static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2220{
2221 struct pnfs_layout_segment *lseg, *tmp;
Trond Myklebusta073dbf2013-03-20 12:34:32 -04002222
2223 /* Matched by references in pnfs_set_layoutcommit */
2224 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2225 list_del_init(&lseg->pls_lc_list);
2226 pnfs_put_lseg(lseg);
2227 }
2228
Trond Myklebust71244d92014-01-13 13:34:36 -05002229 pnfs_clear_layoutcommitting(inode);
Trond Myklebusta073dbf2013-03-20 12:34:32 -04002230}
2231
Peng Tao1b0ae062011-09-22 21:50:12 -04002232void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2233{
Trond Myklebustb9e028f2012-09-18 16:41:18 -04002234 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
Peng Tao1b0ae062011-09-22 21:50:12 -04002235}
2236EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2237
Andy Adamson863a3c62011-03-23 13:27:54 +00002238void
Trond Myklebust67af7612015-03-25 20:40:38 -04002239pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2240 loff_t end_pos)
Andy Adamson863a3c62011-03-23 13:27:54 +00002241{
Fred Isamancd841602012-04-20 14:47:44 -04002242 struct nfs_inode *nfsi = NFS_I(inode);
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04002243 bool mark_as_dirty = false;
Andy Adamson863a3c62011-03-23 13:27:54 +00002244
Fred Isamancd841602012-04-20 14:47:44 -04002245 spin_lock(&inode->i_lock);
Andy Adamson863a3c62011-03-23 13:27:54 +00002246 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
Trond Myklebust29559b12015-03-25 20:10:20 -04002247 nfsi->layout->plh_lwb = end_pos;
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04002248 mark_as_dirty = true;
Andy Adamson863a3c62011-03-23 13:27:54 +00002249 dprintk("%s: Set layoutcommit for inode %lu ",
Fred Isamancd841602012-04-20 14:47:44 -04002250 __func__, inode->i_ino);
Trond Myklebust29559b12015-03-25 20:10:20 -04002251 } else if (end_pos > nfsi->layout->plh_lwb)
2252 nfsi->layout->plh_lwb = end_pos;
Trond Myklebust67af7612015-03-25 20:40:38 -04002253 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
Peng Taoa9bae562011-07-30 20:52:33 -04002254 /* references matched in nfs4_layoutcommit_release */
Trond Myklebust67af7612015-03-25 20:40:38 -04002255 pnfs_get_lseg(lseg);
Peng Taoa9bae562011-07-30 20:52:33 -04002256 }
Fred Isamancd841602012-04-20 14:47:44 -04002257 spin_unlock(&inode->i_lock);
Peng Taoacff58802011-07-30 20:52:31 -04002258 dprintk("%s: lseg %p end_pos %llu\n",
Trond Myklebust67af7612015-03-25 20:40:38 -04002259 __func__, lseg, nfsi->layout->plh_lwb);
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04002260
2261 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2262 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2263 if (mark_as_dirty)
Fred Isamancd841602012-04-20 14:47:44 -04002264 mark_inode_dirty_sync(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00002265}
2266EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2267
Andy Adamsondb29c082011-07-30 20:52:38 -04002268void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2269{
2270 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2271
2272 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2273 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
Trond Myklebusta073dbf2013-03-20 12:34:32 -04002274 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
Andy Adamsondb29c082011-07-30 20:52:38 -04002275}
2276
Andy Adamsonde4b15c2011-03-12 02:58:09 -05002277/*
2278 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2279 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2280 * data to disk to allow the server to recover the data if it crashes.
2281 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2282 * is off, and a COMMIT is sent to a data server, or
2283 * if WRITEs to a data server return NFS_DATA_SYNC.
2284 */
Andy Adamson863a3c62011-03-23 13:27:54 +00002285int
Andy Adamsonef311532011-03-12 02:58:10 -05002286pnfs_layoutcommit_inode(struct inode *inode, bool sync)
Andy Adamson863a3c62011-03-23 13:27:54 +00002287{
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002288 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
Andy Adamson863a3c62011-03-23 13:27:54 +00002289 struct nfs4_layoutcommit_data *data;
2290 struct nfs_inode *nfsi = NFS_I(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00002291 loff_t end_pos;
Trond Myklebust71244d92014-01-13 13:34:36 -05002292 int status;
2293
2294 if (!pnfs_layoutcommit_outstanding(inode))
2295 return 0;
Andy Adamson863a3c62011-03-23 13:27:54 +00002296
2297 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2298
Trond Myklebust71244d92014-01-13 13:34:36 -05002299 status = -EAGAIN;
2300 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2301 if (!sync)
2302 goto out;
NeilBrown74316202014-07-07 15:16:04 +10002303 status = wait_on_bit_lock_action(&nfsi->flags,
Trond Myklebust71244d92014-01-13 13:34:36 -05002304 NFS_INO_LAYOUTCOMMITTING,
2305 nfs_wait_bit_killable,
2306 TASK_KILLABLE);
2307 if (status)
2308 goto out;
2309 }
Andy Adamsonde4b15c2011-03-12 02:58:09 -05002310
Trond Myklebust71244d92014-01-13 13:34:36 -05002311 status = -ENOMEM;
Andy Adamson863a3c62011-03-23 13:27:54 +00002312 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2313 data = kzalloc(sizeof(*data), GFP_NOFS);
Trond Myklebust71244d92014-01-13 13:34:36 -05002314 if (!data)
2315 goto clear_layoutcommitting;
Andy Adamson863a3c62011-03-23 13:27:54 +00002316
Trond Myklebust71244d92014-01-13 13:34:36 -05002317 status = 0;
2318 spin_lock(&inode->i_lock);
2319 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2320 goto out_unlock;
Peng Tao92407e72011-10-23 20:21:17 -07002321
Peng Taoa9bae562011-07-30 20:52:33 -04002322 INIT_LIST_HEAD(&data->lseg_list);
Peng Taoa9bae562011-07-30 20:52:33 -04002323 pnfs_list_write_lseg(inode, &data->lseg_list);
Andy Adamson863a3c62011-03-23 13:27:54 +00002324
Peng Taoacff58802011-07-30 20:52:31 -04002325 end_pos = nfsi->layout->plh_lwb;
Andy Adamson863a3c62011-03-23 13:27:54 +00002326
Trond Myklebustf597c532012-03-04 18:13:56 -05002327 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
Andy Adamson863a3c62011-03-23 13:27:54 +00002328 spin_unlock(&inode->i_lock);
2329
2330 data->args.inode = inode;
Peng Tao9fa40752011-07-30 20:52:32 -04002331 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
Andy Adamson863a3c62011-03-23 13:27:54 +00002332 nfs_fattr_init(&data->fattr);
2333 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2334 data->res.fattr = &data->fattr;
2335 data->args.lastbytewritten = end_pos - 1;
2336 data->res.server = NFS_SERVER(inode);
2337
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002338 if (ld->prepare_layoutcommit) {
2339 status = ld->prepare_layoutcommit(&data->args);
2340 if (status) {
Jeff Layton3471648a2015-07-10 15:58:42 -04002341 put_rpccred(data->cred);
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002342 spin_lock(&inode->i_lock);
Trond Myklebust29559b12015-03-25 20:10:20 -04002343 set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2344 if (end_pos > nfsi->layout->plh_lwb)
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002345 nfsi->layout->plh_lwb = end_pos;
Jeff Layton3471648a2015-07-10 15:58:42 -04002346 goto out_unlock;
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002347 }
2348 }
2349
2350
Andy Adamson863a3c62011-03-23 13:27:54 +00002351 status = nfs4_proc_layoutcommit(data, sync);
2352out:
Peng Tao92407e72011-10-23 20:21:17 -07002353 if (status)
2354 mark_inode_dirty_sync(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00002355 dprintk("<-- %s status %d\n", __func__, status);
2356 return status;
Trond Myklebust71244d92014-01-13 13:34:36 -05002357out_unlock:
2358 spin_unlock(&inode->i_lock);
Peng Tao92407e72011-10-23 20:21:17 -07002359 kfree(data);
Trond Myklebust71244d92014-01-13 13:34:36 -05002360clear_layoutcommitting:
2361 pnfs_clear_layoutcommitting(inode);
Peng Tao92407e72011-10-23 20:21:17 -07002362 goto out;
Andy Adamson863a3c62011-03-23 13:27:54 +00002363}
Peng Tao72cff442014-08-07 10:12:38 +08002364EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
Andy Adamson82be4172012-05-23 05:02:35 -04002365
Trond Myklebust5bb89b42015-03-25 14:14:42 -04002366int
2367pnfs_generic_sync(struct inode *inode, bool datasync)
2368{
2369 return pnfs_layoutcommit_inode(inode, true);
2370}
2371EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2372
Andy Adamson82be4172012-05-23 05:02:35 -04002373struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2374{
2375 struct nfs4_threshold *thp;
2376
2377 thp = kzalloc(sizeof(*thp), GFP_NOFS);
2378 if (!thp) {
2379 dprintk("%s mdsthreshold allocation failed\n", __func__);
2380 return NULL;
2381 }
2382 return thp;
2383}
Peng Tao87334082015-06-23 19:51:57 +08002384
Peng Tao865a7ec2015-06-25 18:19:32 +08002385#if IS_ENABLED(CONFIG_NFS_V4_2)
Peng Tao87334082015-06-23 19:51:57 +08002386int
Trond Myklebustc8ad8892015-08-05 17:31:58 -04002387pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
Peng Tao87334082015-06-23 19:51:57 +08002388{
2389 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2390 struct nfs_server *server = NFS_SERVER(inode);
Peng Tao1bfe3b22015-06-23 19:52:03 +08002391 struct nfs_inode *nfsi = NFS_I(inode);
Peng Tao87334082015-06-23 19:51:57 +08002392 struct nfs42_layoutstat_data *data;
2393 struct pnfs_layout_hdr *hdr;
2394 int status = 0;
2395
2396 if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2397 goto out;
2398
Trond Myklebust6c5a0d82015-06-27 11:45:46 -04002399 if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2400 goto out;
2401
Peng Tao1bfe3b22015-06-23 19:52:03 +08002402 if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2403 goto out;
2404
Peng Tao87334082015-06-23 19:51:57 +08002405 spin_lock(&inode->i_lock);
2406 if (!NFS_I(inode)->layout) {
2407 spin_unlock(&inode->i_lock);
2408 goto out;
2409 }
2410 hdr = NFS_I(inode)->layout;
2411 pnfs_get_layout_hdr(hdr);
2412 spin_unlock(&inode->i_lock);
2413
Trond Myklebustc8ad8892015-08-05 17:31:58 -04002414 data = kzalloc(sizeof(*data), gfp_flags);
Peng Tao87334082015-06-23 19:51:57 +08002415 if (!data) {
2416 status = -ENOMEM;
2417 goto out_put;
2418 }
2419
2420 data->args.fh = NFS_FH(inode);
2421 data->args.inode = inode;
2422 nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2423 status = ld->prepare_layoutstats(&data->args);
2424 if (status)
2425 goto out_free;
2426
2427 status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2428
2429out:
2430 dprintk("%s returns %d\n", __func__, status);
2431 return status;
2432
2433out_free:
2434 kfree(data);
2435out_put:
2436 pnfs_put_layout_hdr(hdr);
Peng Tao1bfe3b22015-06-23 19:52:03 +08002437 smp_mb__before_atomic();
2438 clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2439 smp_mb__after_atomic();
Peng Tao87334082015-06-23 19:51:57 +08002440 goto out;
2441}
2442EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
Peng Tao865a7ec2015-06-25 18:19:32 +08002443#endif
Trond Myklebustbbf58bf2015-08-24 20:39:18 -04002444
2445unsigned int layoutstats_timer;
2446module_param(layoutstats_timer, uint, 0644);
2447EXPORT_SYMBOL_GPL(layoutstats_timer);