blob: 8d9fa4e40e89c7579ae336c45dcd5a29f943c6bf [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 Myklebustb9e028f2012-09-18 16:41:18 -0400255static int
256pnfs_iomode_to_fail_bit(u32 iomode)
257{
258 return iomode == IOMODE_RW ?
259 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
260}
261
262static void
Trond Myklebust3e621212012-09-24 13:07:16 -0400263pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400264{
Trond Myklebust25c75332012-09-18 17:01:12 -0400265 lo->plh_retry_timestamp = jiffies;
Yanchuan Nian39e88fc2013-01-04 20:19:49 +0800266 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
Trond Myklebust3e621212012-09-24 13:07:16 -0400267 atomic_inc(&lo->plh_refcount);
268}
269
270static void
271pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
272{
273 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
274 atomic_dec(&lo->plh_refcount);
275}
276
277static void
278pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
279{
280 struct inode *inode = lo->plh_inode;
Trond Myklebust115ce572012-09-20 21:19:43 -0400281 struct pnfs_layout_range range = {
282 .iomode = iomode,
283 .offset = 0,
284 .length = NFS4_MAX_UINT64,
285 };
286 LIST_HEAD(head);
Trond Myklebust3e621212012-09-24 13:07:16 -0400287
288 spin_lock(&inode->i_lock);
289 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
Trond Myklebust115ce572012-09-20 21:19:43 -0400290 pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
Trond Myklebust3e621212012-09-24 13:07:16 -0400291 spin_unlock(&inode->i_lock);
Trond Myklebust115ce572012-09-20 21:19:43 -0400292 pnfs_free_lseg_list(&head);
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400293 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
294 iomode == IOMODE_RW ? "RW" : "READ");
295}
296
297static bool
298pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
299{
Trond Myklebust25c75332012-09-18 17:01:12 -0400300 unsigned long start, end;
Trond Myklebust3e621212012-09-24 13:07:16 -0400301 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
302
303 if (test_bit(fail_bit, &lo->plh_flags) == 0)
Trond Myklebust25c75332012-09-18 17:01:12 -0400304 return false;
305 end = jiffies;
306 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
307 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
308 /* It is time to retry the failed layoutgets */
Trond Myklebust3e621212012-09-24 13:07:16 -0400309 pnfs_layout_clear_fail_bit(lo, fail_bit);
Trond Myklebust25c75332012-09-18 17:01:12 -0400310 return false;
311 }
312 return true;
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400313}
314
Andy Adamson974cec82010-10-20 00:18:02 -0400315static void
316init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
317{
Fred Isaman566052c2011-01-06 11:36:20 +0000318 INIT_LIST_HEAD(&lseg->pls_list);
Peng Taoa9bae562011-07-30 20:52:33 -0400319 INIT_LIST_HEAD(&lseg->pls_lc_list);
Fred Isaman4541d162011-01-06 11:36:23 +0000320 atomic_set(&lseg->pls_refcount, 1);
321 smp_mb();
322 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
Fred Isaman566052c2011-01-06 11:36:20 +0000323 lseg->pls_layout = lo;
Andy Adamson974cec82010-10-20 00:18:02 -0400324}
325
Trond Myklebust905ca192012-09-20 20:46:49 -0400326static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
Andy Adamson974cec82010-10-20 00:18:02 -0400327{
Fred Isamanb7edfaa2011-01-06 11:36:21 +0000328 struct inode *ino = lseg->pls_layout->plh_inode;
Andy Adamson974cec82010-10-20 00:18:02 -0400329
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400330 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400331}
332
Fred Isamand684d2a2011-03-01 01:34:13 +0000333static void
Trond Myklebust57036a32012-09-20 16:33:30 -0400334pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
335 struct pnfs_layout_segment *lseg)
Andy Adamson974cec82010-10-20 00:18:02 -0400336{
Trond Myklebust57036a32012-09-20 16:33:30 -0400337 struct inode *inode = lo->plh_inode;
Fred Isamand684d2a2011-03-01 01:34:13 +0000338
Benny Halevyd20581a2011-05-22 19:52:03 +0300339 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
Fred Isamand684d2a2011-03-01 01:34:13 +0000340 list_del_init(&lseg->pls_list);
Trond Myklebust8f0d27d2012-09-20 20:57:11 -0400341 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
342 atomic_dec(&lo->plh_refcount);
Trond Myklebust173f77e9c2012-09-21 15:49:42 -0400343 if (list_empty(&lo->plh_segs))
344 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
Fred Isamand684d2a2011-03-01 01:34:13 +0000345 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
346}
347
Fred Isamanbae724e2011-03-01 01:34:15 +0000348void
Trond Myklebust9369a432012-09-18 20:57:08 -0400349pnfs_put_lseg(struct pnfs_layout_segment *lseg)
Fred Isamand684d2a2011-03-01 01:34:13 +0000350{
Trond Myklebust57036a32012-09-20 16:33:30 -0400351 struct pnfs_layout_hdr *lo;
Fred Isamand684d2a2011-03-01 01:34:13 +0000352 struct inode *inode;
353
354 if (!lseg)
355 return;
356
Fred Isaman4541d162011-01-06 11:36:23 +0000357 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
358 atomic_read(&lseg->pls_refcount),
359 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
Trond Myklebust4ef2e4f2015-02-05 17:27:39 -0500360
Trond Myklebust57036a32012-09-20 16:33:30 -0400361 lo = lseg->pls_layout;
362 inode = lo->plh_inode;
Trond Myklebust4ef2e4f2015-02-05 17:27:39 -0500363
Fred Isamand684d2a2011-03-01 01:34:13 +0000364 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
Trond Myklebustfaa4a542015-07-09 18:24:07 +0200365 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
366 spin_unlock(&inode->i_lock);
367 return;
368 }
Trond Myklebust8f0d27d2012-09-20 20:57:11 -0400369 pnfs_get_layout_hdr(lo);
Trond Myklebust4ef2e4f2015-02-05 17:27:39 -0500370 pnfs_layout_remove_lseg(lo, lseg);
371 spin_unlock(&inode->i_lock);
372 pnfs_free_lseg(lseg);
373 pnfs_put_layout_hdr(lo);
Fred Isaman4541d162011-01-06 11:36:23 +0000374 }
Andy Adamson974cec82010-10-20 00:18:02 -0400375}
Trond Myklebust9369a432012-09-18 20:57:08 -0400376EXPORT_SYMBOL_GPL(pnfs_put_lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400377
Trond Myklebust6543f802014-10-08 16:39:12 -0400378static void pnfs_free_lseg_async_work(struct work_struct *work)
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400379{
380 struct pnfs_layout_segment *lseg;
Trond Myklebust6543f802014-10-08 16:39:12 -0400381 struct pnfs_layout_hdr *lo;
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400382
383 lseg = container_of(work, struct pnfs_layout_segment, pls_work);
Trond Myklebust6543f802014-10-08 16:39:12 -0400384 lo = lseg->pls_layout;
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400385
Trond Myklebust6543f802014-10-08 16:39:12 -0400386 pnfs_free_lseg(lseg);
387 pnfs_put_layout_hdr(lo);
388}
389
390static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
391{
392 INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
393 schedule_work(&lseg->pls_work);
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400394}
395
396void
Trond Myklebust6543f802014-10-08 16:39:12 -0400397pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400398{
Trond Myklebust6543f802014-10-08 16:39:12 -0400399 if (!lseg)
400 return;
401
402 assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
403
404 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
405 atomic_read(&lseg->pls_refcount),
406 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
407 if (atomic_dec_and_test(&lseg->pls_refcount)) {
408 struct pnfs_layout_hdr *lo = lseg->pls_layout;
Trond Myklebustfaa4a542015-07-09 18:24:07 +0200409 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
410 return;
Trond Myklebust6543f802014-10-08 16:39:12 -0400411 pnfs_get_layout_hdr(lo);
412 pnfs_layout_remove_lseg(lo, lseg);
413 pnfs_free_lseg_async(lseg);
414 }
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400415}
Trond Myklebust6543f802014-10-08 16:39:12 -0400416EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400417
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400418static u64
Benny Halevyfb3296e2011-05-22 19:47:26 +0300419end_offset(u64 start, u64 len)
Fred Isaman4541d162011-01-06 11:36:23 +0000420{
Benny Halevyfb3296e2011-05-22 19:47:26 +0300421 u64 end;
422
423 end = start + len;
424 return end >= start ? end : NFS4_MAX_UINT64;
425}
426
Benny Halevyfb3296e2011-05-22 19:47:26 +0300427/*
428 * is l2 fully contained in l1?
429 * start1 end1
430 * [----------------------------------)
431 * start2 end2
432 * [----------------)
433 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400434static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400435pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400436 const struct pnfs_layout_range *l2)
Benny Halevyfb3296e2011-05-22 19:47:26 +0300437{
438 u64 start1 = l1->offset;
439 u64 end1 = end_offset(start1, l1->length);
440 u64 start2 = l2->offset;
441 u64 end2 = end_offset(start2, l2->length);
442
443 return (start1 <= start2) && (end1 >= end2);
444}
445
446/*
447 * is l1 and l2 intersecting?
448 * start1 end1
449 * [----------------------------------)
450 * start2 end2
451 * [----------------)
452 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400453static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400454pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400455 const struct pnfs_layout_range *l2)
Benny Halevyfb3296e2011-05-22 19:47:26 +0300456{
457 u64 start1 = l1->offset;
458 u64 end1 = end_offset(start1, l1->length);
459 u64 start2 = l2->offset;
460 u64 end2 = end_offset(start2, l2->length);
461
462 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
463 (end2 == NFS4_MAX_UINT64 || end2 > start1);
464}
465
Fred Isaman4541d162011-01-06 11:36:23 +0000466static bool
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400467should_free_lseg(const struct pnfs_layout_range *lseg_range,
468 const struct pnfs_layout_range *recall_range)
Fred Isaman4541d162011-01-06 11:36:23 +0000469{
Benny Halevy778b5502011-05-22 19:48:02 +0300470 return (recall_range->iomode == IOMODE_ANY ||
471 lseg_range->iomode == recall_range->iomode) &&
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400472 pnfs_lseg_range_intersecting(lseg_range, recall_range);
Fred Isaman4541d162011-01-06 11:36:23 +0000473}
474
Trond Myklebust24956802013-03-20 13:03:00 -0400475static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
476 struct list_head *tmp_list)
477{
478 if (!atomic_dec_and_test(&lseg->pls_refcount))
479 return false;
480 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
481 list_add(&lseg->pls_list, tmp_list);
482 return true;
483}
484
Fred Isaman4541d162011-01-06 11:36:23 +0000485/* Returns 1 if lseg is removed from list, 0 otherwise */
486static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
487 struct list_head *tmp_list)
488{
489 int rv = 0;
490
491 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
492 /* Remove the reference keeping the lseg in the
493 * list. It will now be removed when all
494 * outstanding io is finished.
495 */
Fred Isamand684d2a2011-03-01 01:34:13 +0000496 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
497 atomic_read(&lseg->pls_refcount));
Trond Myklebust24956802013-03-20 13:03:00 -0400498 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
Fred Isamand684d2a2011-03-01 01:34:13 +0000499 rv = 1;
Fred Isaman4541d162011-01-06 11:36:23 +0000500 }
501 return rv;
502}
503
504/* Returns count of number of matching invalid lsegs remaining in list
505 * after call.
506 */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000507int
Trond Myklebust49a85062012-09-18 20:43:31 -0400508pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
Fred Isaman4541d162011-01-06 11:36:23 +0000509 struct list_head *tmp_list,
Trond Myklebust506c0d62016-01-04 13:04:47 -0500510 const struct pnfs_layout_range *recall_range)
Andy Adamson974cec82010-10-20 00:18:02 -0400511{
512 struct pnfs_layout_segment *lseg, *next;
Trond Myklebust71b39852016-01-04 12:41:15 -0500513 int remaining = 0;
Andy Adamson974cec82010-10-20 00:18:02 -0400514
515 dprintk("%s:Begin lo %p\n", __func__, lo);
516
Trond Myklebust8006bfb2012-09-21 14:48:04 -0400517 if (list_empty(&lo->plh_segs))
Fred Isaman385117222011-02-03 18:28:50 +0000518 return 0;
Fred Isaman4541d162011-01-06 11:36:23 +0000519 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
Benny Halevy778b5502011-05-22 19:48:02 +0300520 if (!recall_range ||
521 should_free_lseg(&lseg->pls_range, recall_range)) {
Fred Isaman4541d162011-01-06 11:36:23 +0000522 dprintk("%s: freeing lseg %p iomode %d "
523 "offset %llu length %llu\n", __func__,
524 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
525 lseg->pls_range.length);
Trond Myklebust71b39852016-01-04 12:41:15 -0500526 if (!mark_lseg_invalid(lseg, tmp_list))
527 remaining++;
Fred Isaman4541d162011-01-06 11:36:23 +0000528 }
Trond Myklebust71b39852016-01-04 12:41:15 -0500529 dprintk("%s:Return %i\n", __func__, remaining);
530 return remaining;
Andy Adamson974cec82010-10-20 00:18:02 -0400531}
532
Fred Isamanf49f9ba2011-02-03 18:28:52 +0000533/* note free_me must contain lsegs from a single layout_hdr */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000534void
Fred Isaman4541d162011-01-06 11:36:23 +0000535pnfs_free_lseg_list(struct list_head *free_me)
Andy Adamson974cec82010-10-20 00:18:02 -0400536{
Fred Isaman4541d162011-01-06 11:36:23 +0000537 struct pnfs_layout_segment *lseg, *tmp;
Andy Adamson974cec82010-10-20 00:18:02 -0400538
Fred Isamanf49f9ba2011-02-03 18:28:52 +0000539 if (list_empty(free_me))
540 return;
541
Fred Isaman4541d162011-01-06 11:36:23 +0000542 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
Fred Isaman566052c2011-01-06 11:36:20 +0000543 list_del(&lseg->pls_list);
Trond Myklebust905ca192012-09-20 20:46:49 -0400544 pnfs_free_lseg(lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400545 }
546}
547
Benny Halevye5e94012010-10-20 00:18:01 -0400548void
549pnfs_destroy_layout(struct nfs_inode *nfsi)
550{
551 struct pnfs_layout_hdr *lo;
Andy Adamson974cec82010-10-20 00:18:02 -0400552 LIST_HEAD(tmp_list);
Benny Halevye5e94012010-10-20 00:18:01 -0400553
554 spin_lock(&nfsi->vfs_inode.i_lock);
555 lo = nfsi->layout;
556 if (lo) {
Fred Isaman385117222011-02-03 18:28:50 +0000557 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
Trond Myklebust49a85062012-09-18 20:43:31 -0400558 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
Trond Myklebust3e621212012-09-24 13:07:16 -0400559 pnfs_get_layout_hdr(lo);
560 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
561 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
562 spin_unlock(&nfsi->vfs_inode.i_lock);
563 pnfs_free_lseg_list(&tmp_list);
564 pnfs_put_layout_hdr(lo);
565 } else
566 spin_unlock(&nfsi->vfs_inode.i_lock);
Benny Halevye5e94012010-10-20 00:18:01 -0400567}
Andy Adamson041245c2012-04-27 17:53:53 -0400568EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
Benny Halevye5e94012010-10-20 00:18:01 -0400569
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500570static bool
571pnfs_layout_add_bulk_destroy_list(struct inode *inode,
572 struct list_head *layout_list)
573{
574 struct pnfs_layout_hdr *lo;
575 bool ret = false;
576
577 spin_lock(&inode->i_lock);
578 lo = NFS_I(inode)->layout;
579 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
580 pnfs_get_layout_hdr(lo);
581 list_add(&lo->plh_bulk_destroy, layout_list);
582 ret = true;
583 }
584 spin_unlock(&inode->i_lock);
585 return ret;
586}
587
588/* Caller must hold rcu_read_lock and clp->cl_lock */
589static int
590pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
591 struct nfs_server *server,
592 struct list_head *layout_list)
593{
594 struct pnfs_layout_hdr *lo, *next;
595 struct inode *inode;
596
597 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
598 inode = igrab(lo->plh_inode);
599 if (inode == NULL)
600 continue;
601 list_del_init(&lo->plh_layouts);
602 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
603 continue;
604 rcu_read_unlock();
605 spin_unlock(&clp->cl_lock);
606 iput(inode);
607 spin_lock(&clp->cl_lock);
608 rcu_read_lock();
609 return -EAGAIN;
610 }
611 return 0;
612}
613
614static int
615pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
616 bool is_bulk_recall)
617{
618 struct pnfs_layout_hdr *lo;
619 struct inode *inode;
620 struct pnfs_layout_range range = {
621 .iomode = IOMODE_ANY,
622 .offset = 0,
623 .length = NFS4_MAX_UINT64,
624 };
625 LIST_HEAD(lseg_list);
626 int ret = 0;
627
628 while (!list_empty(layout_list)) {
629 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
630 plh_bulk_destroy);
631 dprintk("%s freeing layout for inode %lu\n", __func__,
632 lo->plh_inode->i_ino);
633 inode = lo->plh_inode;
Christoph Hellwig7c5d1872014-09-10 08:23:29 -0700634
635 pnfs_layoutcommit_inode(inode, false);
636
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500637 spin_lock(&inode->i_lock);
638 list_del_init(&lo->plh_bulk_destroy);
639 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
640 if (is_bulk_recall)
641 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
642 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
643 ret = -EAGAIN;
644 spin_unlock(&inode->i_lock);
645 pnfs_free_lseg_list(&lseg_list);
Trond Myklebustb20135d2015-12-31 09:28:06 -0500646 /* Free all lsegs that are attached to commit buckets */
647 nfs_commit_inode(inode, 0);
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500648 pnfs_put_layout_hdr(lo);
649 iput(inode);
650 }
651 return ret;
652}
653
654int
655pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
656 struct nfs_fsid *fsid,
657 bool is_recall)
658{
659 struct nfs_server *server;
660 LIST_HEAD(layout_list);
661
662 spin_lock(&clp->cl_lock);
663 rcu_read_lock();
664restart:
665 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
666 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
667 continue;
668 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
669 server,
670 &layout_list) != 0)
671 goto restart;
672 }
673 rcu_read_unlock();
674 spin_unlock(&clp->cl_lock);
675
676 if (list_empty(&layout_list))
677 return 0;
678 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
679}
680
681int
682pnfs_destroy_layouts_byclid(struct nfs_client *clp,
683 bool is_recall)
684{
685 struct nfs_server *server;
686 LIST_HEAD(layout_list);
687
688 spin_lock(&clp->cl_lock);
689 rcu_read_lock();
690restart:
691 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
692 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
693 server,
694 &layout_list) != 0)
695 goto restart;
696 }
697 rcu_read_unlock();
698 spin_unlock(&clp->cl_lock);
699
700 if (list_empty(&layout_list))
701 return 0;
702 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
703}
704
Andy Adamson974cec82010-10-20 00:18:02 -0400705/*
706 * Called by the state manger to remove all layouts established under an
707 * expired lease.
708 */
709void
710pnfs_destroy_all_layouts(struct nfs_client *clp)
711{
Andy Adamsonc47abcf2011-06-15 17:52:40 -0400712 nfs4_deviceid_mark_client_invalid(clp);
713 nfs4_deviceid_purge_client(clp);
714
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500715 pnfs_destroy_layouts_byclid(clp, false);
Andy Adamson974cec82010-10-20 00:18:02 -0400716}
717
Trond Myklebust5a655032012-10-02 16:47:14 -0700718/*
719 * Compare 2 layout stateid sequence ids, to see which is newer,
720 * taking into account wraparound issues.
721 */
722static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
723{
Trond Myklebust2c64c572014-02-12 09:36:59 -0500724 return (s32)(s1 - s2) > 0;
Trond Myklebust5a655032012-10-02 16:47:14 -0700725}
726
Fred Isamanfd6002e2011-01-06 11:36:22 +0000727/* update lo->plh_stateid with new if is more recent */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000728void
729pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
730 bool update_barrier)
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400731{
Trond Myklebust22aaf712012-10-04 16:32:22 -0700732 u32 oldseq, newseq, new_barrier;
733 int empty = list_empty(&lo->plh_segs);
Andy Adamson974cec82010-10-20 00:18:02 -0400734
Trond Myklebust2d2f24a2012-03-04 18:13:57 -0500735 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
736 newseq = be32_to_cpu(new->seqid);
Trond Myklebust22aaf712012-10-04 16:32:22 -0700737 if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
Trond Myklebustf597c532012-03-04 18:13:56 -0500738 nfs4_stateid_copy(&lo->plh_stateid, new);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000739 if (update_barrier) {
Trond Myklebust22aaf712012-10-04 16:32:22 -0700740 new_barrier = be32_to_cpu(new->seqid);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000741 } else {
742 /* Because of wraparound, we want to keep the barrier
Trond Myklebust22aaf712012-10-04 16:32:22 -0700743 * "close" to the current seqids.
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000744 */
Trond Myklebust22aaf712012-10-04 16:32:22 -0700745 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000746 }
Trond Myklebust22aaf712012-10-04 16:32:22 -0700747 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
748 lo->plh_barrier = new_barrier;
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000749 }
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400750}
751
Trond Myklebust19c54ab2012-10-05 16:56:58 -0700752static bool
753pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
754 const nfs4_stateid *stateid)
755{
756 u32 seqid = be32_to_cpu(stateid->seqid);
757
758 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
759}
760
Fred Isamancf7d63f2011-01-06 11:36:25 +0000761/* lget is set to 1 if called from inside send_layoutget call chain */
762static bool
Trond Myklebuste1c06f82015-08-04 16:40:08 -0400763pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
Fred Isamancf7d63f2011-01-06 11:36:25 +0000764{
Fred Isamanf7e89172011-01-06 11:36:32 +0000765 return lo->plh_block_lgets ||
Trond Myklebuste1c06f82015-08-04 16:40:08 -0400766 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
Fred Isamancf7d63f2011-01-06 11:36:25 +0000767}
768
Fred Isamanfd6002e2011-01-06 11:36:22 +0000769int
770pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
Trond Myklebust506c0d62016-01-04 13:04:47 -0500771 const struct pnfs_layout_range *range,
Fred Isamanfd6002e2011-01-06 11:36:22 +0000772 struct nfs4_state *open_state)
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400773{
Fred Isamanfd6002e2011-01-06 11:36:22 +0000774 int status = 0;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400775
776 dprintk("--> %s\n", __func__);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000777 spin_lock(&lo->plh_inode->i_lock);
Trond Myklebuste1c06f82015-08-04 16:40:08 -0400778 if (pnfs_layoutgets_blocked(lo)) {
Fred Isamancf7d63f2011-01-06 11:36:25 +0000779 status = -EAGAIN;
Trond Myklebust5d422302013-03-14 16:57:48 -0400780 } else if (!nfs4_valid_open_stateid(open_state)) {
781 status = -EBADF;
Christoph Hellwig47abade2014-08-21 11:09:22 -0500782 } else if (list_empty(&lo->plh_segs) ||
783 test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
Fred Isamanfd6002e2011-01-06 11:36:22 +0000784 int seq;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400785
Fred Isamanfd6002e2011-01-06 11:36:22 +0000786 do {
787 seq = read_seqbegin(&open_state->seqlock);
Trond Myklebustf597c532012-03-04 18:13:56 -0500788 nfs4_stateid_copy(dst, &open_state->stateid);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000789 } while (read_seqretry(&open_state->seqlock, seq));
790 } else
Trond Myklebustf597c532012-03-04 18:13:56 -0500791 nfs4_stateid_copy(dst, &lo->plh_stateid);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000792 spin_unlock(&lo->plh_inode->i_lock);
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400793 dprintk("<-- %s\n", __func__);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000794 return status;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400795}
796
797/*
798* Get layout from server.
799* for now, assume that whole file layouts are requested.
800* arg->offset: 0
801* arg->length: all ones
802*/
Benny Halevye5e94012010-10-20 00:18:01 -0400803static struct pnfs_layout_segment *
804send_layoutget(struct pnfs_layout_hdr *lo,
805 struct nfs_open_context *ctx,
Trond Myklebuste144e532016-01-04 12:52:53 -0500806 const struct pnfs_layout_range *range,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400807 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -0400808{
Fred Isamanb7edfaa2011-01-06 11:36:21 +0000809 struct inode *ino = lo->plh_inode;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400810 struct nfs_server *server = NFS_SERVER(ino);
811 struct nfs4_layoutget *lgp;
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -0400812 struct pnfs_layout_segment *lseg;
Trond Myklebust2d89a1d2015-08-31 02:05:47 -0700813 loff_t i_size;
Benny Halevye5e94012010-10-20 00:18:01 -0400814
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400815 dprintk("--> %s\n", __func__);
816
Jeff Layton4f2e9dc2015-11-25 13:43:14 -0500817 /*
818 * Synchronously retrieve layout information from server and
819 * store in lseg. If we race with a concurrent seqid morphing
820 * op, then re-send the LAYOUTGET.
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400821 */
Jeff Layton4f2e9dc2015-11-25 13:43:14 -0500822 do {
823 lgp = kzalloc(sizeof(*lgp), gfp_flags);
824 if (lgp == NULL)
825 return NULL;
826
827 i_size = i_size_read(ino);
828
829 lgp->args.minlength = PAGE_CACHE_SIZE;
830 if (lgp->args.minlength > range->length)
831 lgp->args.minlength = range->length;
832 if (range->iomode == IOMODE_READ) {
833 if (range->offset >= i_size)
834 lgp->args.minlength = 0;
835 else if (i_size - range->offset < lgp->args.minlength)
836 lgp->args.minlength = i_size - range->offset;
837 }
838 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
Trond Myklebuste144e532016-01-04 12:52:53 -0500839 pnfs_copy_range(&lgp->args.range, range);
Jeff Layton4f2e9dc2015-11-25 13:43:14 -0500840 lgp->args.type = server->pnfs_curr_ld->id;
841 lgp->args.inode = ino;
842 lgp->args.ctx = get_nfs_open_context(ctx);
843 lgp->gfp_flags = gfp_flags;
844 lgp->cred = lo->plh_lc_cred;
845
846 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
847 } while (lseg == ERR_PTR(-EAGAIN));
848
Peng Tao0bcbf032015-12-05 15:57:31 +0800849 if (IS_ERR(lseg) && !nfs_error_is_fatal(PTR_ERR(lseg)))
850 lseg = NULL;
851 else
Tom Haynesd67ae822014-12-11 17:02:04 -0500852 pnfs_layout_clear_fail_bit(lo,
853 pnfs_iomode_to_fail_bit(range->iomode));
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400854
Andy Adamson974cec82010-10-20 00:18:02 -0400855 return lseg;
856}
857
Trond Myklebust24956802013-03-20 13:03:00 -0400858static void pnfs_clear_layoutcommit(struct inode *inode,
859 struct list_head *head)
860{
861 struct nfs_inode *nfsi = NFS_I(inode);
862 struct pnfs_layout_segment *lseg, *tmp;
863
864 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
865 return;
866 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
867 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
868 continue;
869 pnfs_lseg_dec_and_remove_zero(lseg, head);
870 }
871}
872
Tom Haynesd67ae822014-12-11 17:02:04 -0500873void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
874{
875 clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
876 smp_mb__after_atomic();
877 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
Trond Myklebust7f273922015-07-09 18:40:01 +0200878 rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
Tom Haynesd67ae822014-12-11 17:02:04 -0500879}
880
Trond Myklebust13c13a62016-01-26 23:12:11 -0500881static bool
882pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
883{
884 if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
885 return false;
886 lo->plh_return_iomode = 0;
887 pnfs_get_layout_hdr(lo);
Trond Myklebust2370abd2016-01-27 20:32:50 -0500888 clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
Trond Myklebust13c13a62016-01-26 23:12:11 -0500889 return true;
890}
891
Peng Taof40eb5d2014-09-06 00:53:22 +0800892static int
Trond Myklebusted429d62016-01-04 12:30:55 -0500893pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
Peng Tao6c166052014-11-17 09:30:40 +0800894 enum pnfs_iomode iomode, bool sync)
Peng Taof40eb5d2014-09-06 00:53:22 +0800895{
896 struct inode *ino = lo->plh_inode;
897 struct nfs4_layoutreturn *lrp;
898 int status = 0;
899
Trond Myklebuste4af4402015-02-05 17:05:08 -0500900 lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
Peng Taof40eb5d2014-09-06 00:53:22 +0800901 if (unlikely(lrp == NULL)) {
902 status = -ENOMEM;
903 spin_lock(&ino->i_lock);
Tom Haynesd67ae822014-12-11 17:02:04 -0500904 pnfs_clear_layoutreturn_waitbit(lo);
Peng Taof40eb5d2014-09-06 00:53:22 +0800905 spin_unlock(&ino->i_lock);
906 pnfs_put_layout_hdr(lo);
907 goto out;
908 }
909
Trond Myklebusted429d62016-01-04 12:30:55 -0500910 nfs4_stateid_copy(&lrp->args.stateid, stateid);
Peng Taof40eb5d2014-09-06 00:53:22 +0800911 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
912 lrp->args.inode = ino;
Peng Tao15eb67c2014-11-17 09:30:36 +0800913 lrp->args.range.iomode = iomode;
914 lrp->args.range.offset = 0;
915 lrp->args.range.length = NFS4_MAX_UINT64;
Peng Taof40eb5d2014-09-06 00:53:22 +0800916 lrp->args.layout = lo;
917 lrp->clp = NFS_SERVER(ino)->nfs_client;
918 lrp->cred = lo->plh_lc_cred;
919
Peng Tao6c166052014-11-17 09:30:40 +0800920 status = nfs4_proc_layoutreturn(lrp, sync);
Peng Taof40eb5d2014-09-06 00:53:22 +0800921out:
922 dprintk("<-- %s status: %d\n", __func__, status);
923 return status;
924}
925
Trond Myklebust13c13a62016-01-26 23:12:11 -0500926/* Return true if layoutreturn is needed */
927static bool
928pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
929{
930 struct pnfs_layout_segment *s;
931
Trond Myklebust2370abd2016-01-27 20:32:50 -0500932 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
Trond Myklebust13c13a62016-01-26 23:12:11 -0500933 return false;
934
935 /* Defer layoutreturn until all lsegs are done */
936 list_for_each_entry(s, &lo->plh_segs, pls_list) {
937 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
938 return false;
939 }
940
941 return true;
942}
943
944static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
945{
946 struct inode *inode= lo->plh_inode;
947
Trond Myklebust2370abd2016-01-27 20:32:50 -0500948 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
Trond Myklebust13c13a62016-01-26 23:12:11 -0500949 return;
950 spin_lock(&inode->i_lock);
951 if (pnfs_layout_need_return(lo)) {
952 nfs4_stateid stateid;
953 enum pnfs_iomode iomode;
954 bool send;
955
956 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
957 iomode = lo->plh_return_iomode;
958 send = pnfs_prepare_layoutreturn(lo);
959 spin_unlock(&inode->i_lock);
960 if (send) {
961 /* Send an async layoutreturn so we dont deadlock */
962 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
963 }
964 } else
965 spin_unlock(&inode->i_lock);
966}
967
Andy Adamson293b3b02012-06-20 15:03:34 -0400968/*
969 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
970 * when the layout segment list is empty.
971 *
972 * Note that a pnfs_layout_hdr can exist with an empty layout segment
973 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
974 * deviceid is marked invalid.
975 */
Benny Halevycbe82602011-05-22 19:52:37 +0300976int
977_pnfs_return_layout(struct inode *ino)
978{
979 struct pnfs_layout_hdr *lo = NULL;
980 struct nfs_inode *nfsi = NFS_I(ino);
981 LIST_HEAD(tmp_list);
Benny Halevycbe82602011-05-22 19:52:37 +0300982 nfs4_stateid stateid;
Andy Adamson293b3b02012-06-20 15:03:34 -0400983 int status = 0, empty;
Trond Myklebust7f273922015-07-09 18:40:01 +0200984 bool send;
Benny Halevycbe82602011-05-22 19:52:37 +0300985
Andy Adamson366d5052012-06-20 15:03:33 -0400986 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
Benny Halevycbe82602011-05-22 19:52:37 +0300987
988 spin_lock(&ino->i_lock);
989 lo = nfsi->layout;
Trond Myklebuste5929f32012-09-21 16:37:02 -0400990 if (!lo) {
Benny Halevycbe82602011-05-22 19:52:37 +0300991 spin_unlock(&ino->i_lock);
Andy Adamson293b3b02012-06-20 15:03:34 -0400992 dprintk("NFS: %s no layout to return\n", __func__);
993 goto out;
Benny Halevycbe82602011-05-22 19:52:37 +0300994 }
Trond Myklebust50f563e2016-01-04 12:22:46 -0500995 nfs4_stateid_copy(&stateid, &nfsi->layout->plh_stateid);
Benny Halevycbe82602011-05-22 19:52:37 +0300996 /* Reference matched in nfs4_layoutreturn_release */
Trond Myklebust70c3bd22012-09-18 20:51:13 -0400997 pnfs_get_layout_hdr(lo);
Andy Adamson293b3b02012-06-20 15:03:34 -0400998 empty = list_empty(&lo->plh_segs);
Trond Myklebust24956802013-03-20 13:03:00 -0400999 pnfs_clear_layoutcommit(ino, &tmp_list);
Trond Myklebust49a85062012-09-18 20:43:31 -04001000 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
Christoph Hellwigc88953d2014-09-10 08:23:31 -07001001
1002 if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1003 struct pnfs_layout_range range = {
1004 .iomode = IOMODE_ANY,
1005 .offset = 0,
1006 .length = NFS4_MAX_UINT64,
1007 };
1008 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1009 }
1010
Andy Adamson293b3b02012-06-20 15:03:34 -04001011 /* Don't send a LAYOUTRETURN if list was initially empty */
1012 if (empty) {
1013 spin_unlock(&ino->i_lock);
Andy Adamson293b3b02012-06-20 15:03:34 -04001014 dprintk("NFS: %s no layout segments to return\n", __func__);
Trond Myklebust7f273922015-07-09 18:40:01 +02001015 goto out_put_layout_hdr;
Andy Adamson293b3b02012-06-20 15:03:34 -04001016 }
Christoph Hellwig47abade2014-08-21 11:09:22 -05001017
1018 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
Trond Myklebust7f273922015-07-09 18:40:01 +02001019 send = pnfs_prepare_layoutreturn(lo);
Benny Halevycbe82602011-05-22 19:52:37 +03001020 spin_unlock(&ino->i_lock);
1021 pnfs_free_lseg_list(&tmp_list);
Trond Myklebust7f273922015-07-09 18:40:01 +02001022 if (send)
Trond Myklebusted429d62016-01-04 12:30:55 -05001023 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
Trond Myklebust7f273922015-07-09 18:40:01 +02001024out_put_layout_hdr:
1025 pnfs_put_layout_hdr(lo);
Benny Halevycbe82602011-05-22 19:52:37 +03001026out:
1027 dprintk("<-- %s status: %d\n", __func__, status);
1028 return status;
1029}
Andy Adamson0a57cda2012-04-27 17:53:50 -04001030EXPORT_SYMBOL_GPL(_pnfs_return_layout);
Benny Halevycbe82602011-05-22 19:52:37 +03001031
Trond Myklebust24028672013-03-20 13:23:33 -04001032int
1033pnfs_commit_and_return_layout(struct inode *inode)
1034{
1035 struct pnfs_layout_hdr *lo;
1036 int ret;
1037
1038 spin_lock(&inode->i_lock);
1039 lo = NFS_I(inode)->layout;
1040 if (lo == NULL) {
1041 spin_unlock(&inode->i_lock);
1042 return 0;
1043 }
1044 pnfs_get_layout_hdr(lo);
1045 /* Block new layoutgets and read/write to ds */
1046 lo->plh_block_lgets++;
1047 spin_unlock(&inode->i_lock);
1048 filemap_fdatawait(inode->i_mapping);
1049 ret = pnfs_layoutcommit_inode(inode, true);
1050 if (ret == 0)
1051 ret = _pnfs_return_layout(inode);
1052 spin_lock(&inode->i_lock);
1053 lo->plh_block_lgets--;
1054 spin_unlock(&inode->i_lock);
1055 pnfs_put_layout_hdr(lo);
1056 return ret;
1057}
1058
Fred Isamanf7e89172011-01-06 11:36:32 +00001059bool pnfs_roc(struct inode *ino)
1060{
Trond Myklebust40dd4b72015-01-24 13:54:37 -05001061 struct nfs_inode *nfsi = NFS_I(ino);
1062 struct nfs_open_context *ctx;
1063 struct nfs4_state *state;
Fred Isamanf7e89172011-01-06 11:36:32 +00001064 struct pnfs_layout_hdr *lo;
1065 struct pnfs_layout_segment *lseg, *tmp;
Peng Tao193e3aa2014-11-17 09:30:41 +08001066 nfs4_stateid stateid;
Fred Isamanf7e89172011-01-06 11:36:32 +00001067 LIST_HEAD(tmp_list);
Peng Taoe755d632015-08-19 13:49:19 +08001068 bool found = false, layoutreturn = false, roc = false;
Fred Isamanf7e89172011-01-06 11:36:32 +00001069
1070 spin_lock(&ino->i_lock);
Trond Myklebust40dd4b72015-01-24 13:54:37 -05001071 lo = nfsi->layout;
Peng Tao39761432015-08-21 12:49:44 +08001072 if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
Trond Myklebust40dd4b72015-01-24 13:54:37 -05001073 goto out_noroc;
1074
Peng Taoe755d632015-08-19 13:49:19 +08001075 /* no roc if we hold a delegation */
Trond Myklebust40dd4b72015-01-24 13:54:37 -05001076 if (nfs4_check_delegation(ino, FMODE_READ))
1077 goto out_noroc;
1078
1079 list_for_each_entry(ctx, &nfsi->open_files, list) {
1080 state = ctx->state;
1081 /* Don't return layout if there is open file state */
1082 if (state != NULL && state->state != 0)
1083 goto out_noroc;
1084 }
1085
Trond Myklebust50f563e2016-01-04 12:22:46 -05001086 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
Peng Taoe755d632015-08-19 13:49:19 +08001087 /* always send layoutreturn if being marked so */
Trond Myklebust2370abd2016-01-27 20:32:50 -05001088 if (test_and_clear_bit(NFS_LAYOUT_RETURN_REQUESTED,
Peng Taoe755d632015-08-19 13:49:19 +08001089 &lo->plh_flags))
1090 layoutreturn = pnfs_prepare_layoutreturn(lo);
1091
Fred Isamanf7e89172011-01-06 11:36:32 +00001092 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
Peng Taoe755d632015-08-19 13:49:19 +08001093 /* If we are sending layoutreturn, invalidate all valid lsegs */
1094 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
Fred Isamanf7e89172011-01-06 11:36:32 +00001095 mark_lseg_invalid(lseg, &tmp_list);
1096 found = true;
1097 }
Peng Tao500d7012015-09-22 11:35:22 +08001098 /* ROC in two conditions:
Peng Taoe755d632015-08-19 13:49:19 +08001099 * 1. there are ROC lsegs
1100 * 2. we don't send layoutreturn
Peng Taoe755d632015-08-19 13:49:19 +08001101 */
Peng Tao500d7012015-09-22 11:35:22 +08001102 if (found && !layoutreturn) {
1103 /* lo ref dropped in pnfs_roc_release() */
1104 pnfs_get_layout_hdr(lo);
Peng Taoe755d632015-08-19 13:49:19 +08001105 roc = true;
Peng Tao500d7012015-09-22 11:35:22 +08001106 }
Peng Taoe755d632015-08-19 13:49:19 +08001107
1108out_noroc:
Fred Isamanf7e89172011-01-06 11:36:32 +00001109 spin_unlock(&ino->i_lock);
1110 pnfs_free_lseg_list(&tmp_list);
Trond Myklebust71401712015-03-25 12:36:13 -04001111 pnfs_layoutcommit_inode(ino, true);
Peng Taoe755d632015-08-19 13:49:19 +08001112 if (layoutreturn)
Trond Myklebusted429d62016-01-04 12:30:55 -05001113 pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
Peng Taoe755d632015-08-19 13:49:19 +08001114 return roc;
Fred Isamanf7e89172011-01-06 11:36:32 +00001115}
1116
1117void pnfs_roc_release(struct inode *ino)
1118{
1119 struct pnfs_layout_hdr *lo;
1120
1121 spin_lock(&ino->i_lock);
1122 lo = NFS_I(ino)->layout;
Trond Myklebust5c4a79f2015-08-04 15:57:13 -04001123 pnfs_clear_layoutreturn_waitbit(lo);
Trond Myklebust6622c3e2012-09-20 17:23:11 -04001124 if (atomic_dec_and_test(&lo->plh_refcount)) {
1125 pnfs_detach_layout_hdr(lo);
1126 spin_unlock(&ino->i_lock);
1127 pnfs_free_layout_hdr(lo);
1128 } else
1129 spin_unlock(&ino->i_lock);
Fred Isamanf7e89172011-01-06 11:36:32 +00001130}
1131
1132void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1133{
1134 struct pnfs_layout_hdr *lo;
1135
1136 spin_lock(&ino->i_lock);
1137 lo = NFS_I(ino)->layout;
Trond Myklebust0654cc72015-12-28 11:48:14 -05001138 pnfs_mark_layout_returned_if_empty(lo);
Trond Myklebust0f35ad62012-10-04 16:28:17 -07001139 if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
Fred Isamanf7e89172011-01-06 11:36:32 +00001140 lo->plh_barrier = barrier;
1141 spin_unlock(&ino->i_lock);
Trond Myklebust6a463beb2015-08-20 15:40:47 -05001142 trace_nfs4_layoutreturn_on_close(ino, 0);
Fred Isamanf7e89172011-01-06 11:36:32 +00001143}
1144
Trond Myklebust4ff376f2015-08-18 23:23:21 -05001145void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
Fred Isamanf7e89172011-01-06 11:36:32 +00001146{
1147 struct nfs_inode *nfsi = NFS_I(ino);
Trond Myklebust7fdab062012-09-20 20:15:57 -04001148 struct pnfs_layout_hdr *lo;
Trond Myklebust7fdab062012-09-20 20:15:57 -04001149 u32 current_seqid;
Fred Isamanf7e89172011-01-06 11:36:32 +00001150
1151 spin_lock(&ino->i_lock);
Trond Myklebust7fdab062012-09-20 20:15:57 -04001152 lo = nfsi->layout;
1153 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
Fred Isamanf7e89172011-01-06 11:36:32 +00001154
Trond Myklebust7fdab062012-09-20 20:15:57 -04001155 /* Since close does not return a layout stateid for use as
1156 * a barrier, we choose the worst-case barrier.
1157 */
1158 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
Fred Isamanf7e89172011-01-06 11:36:32 +00001159 spin_unlock(&ino->i_lock);
Fred Isamanf7e89172011-01-06 11:36:32 +00001160}
1161
Peng Tao500d7012015-09-22 11:35:22 +08001162bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1163{
1164 struct nfs_inode *nfsi = NFS_I(ino);
1165 struct pnfs_layout_hdr *lo;
1166 bool sleep = false;
1167
1168 /* we might not have grabbed lo reference. so need to check under
1169 * i_lock */
1170 spin_lock(&ino->i_lock);
1171 lo = nfsi->layout;
1172 if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
1173 sleep = true;
1174 spin_unlock(&ino->i_lock);
1175
1176 if (sleep)
1177 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1178
1179 return sleep;
1180}
1181
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001182/*
1183 * Compare two layout segments for sorting into layout cache.
1184 * We want to preferentially return RW over RO layouts, so ensure those
1185 * are seen first.
1186 */
1187static s64
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001188pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001189 const struct pnfs_layout_range *l2)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001190{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001191 s64 d;
1192
1193 /* high offset > low offset */
1194 d = l1->offset - l2->offset;
1195 if (d)
1196 return d;
1197
1198 /* short length > long length */
1199 d = l2->length - l1->length;
1200 if (d)
1201 return d;
1202
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001203 /* read > read/write */
Benny Halevyfb3296e2011-05-22 19:47:26 +03001204 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001205}
1206
Trond Myklebust03772d22015-08-25 08:54:17 -04001207static bool
1208pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1209 const struct pnfs_layout_range *l2)
Andy Adamson974cec82010-10-20 00:18:02 -04001210{
Trond Myklebust03772d22015-08-25 08:54:17 -04001211 return pnfs_lseg_range_cmp(l1, l2) > 0;
1212}
1213
1214static bool
1215pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1216 struct pnfs_layout_segment *old)
1217{
1218 return false;
1219}
1220
1221void
1222pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1223 struct pnfs_layout_segment *lseg,
1224 bool (*is_after)(const struct pnfs_layout_range *,
1225 const struct pnfs_layout_range *),
1226 bool (*do_merge)(struct pnfs_layout_segment *,
1227 struct pnfs_layout_segment *),
1228 struct list_head *free_me)
1229{
1230 struct pnfs_layout_segment *lp, *tmp;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001231
Andy Adamson974cec82010-10-20 00:18:02 -04001232 dprintk("%s:Begin\n", __func__);
1233
Trond Myklebust03772d22015-08-25 08:54:17 -04001234 list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1235 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1236 continue;
1237 if (do_merge(lseg, lp)) {
1238 mark_lseg_invalid(lp, free_me);
1239 continue;
1240 }
1241 if (is_after(&lseg->pls_range, &lp->pls_range))
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001242 continue;
Fred Isaman566052c2011-01-06 11:36:20 +00001243 list_add_tail(&lseg->pls_list, &lp->pls_list);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001244 dprintk("%s: inserted lseg %p "
1245 "iomode %d offset %llu length %llu before "
1246 "lp %p iomode %d offset %llu length %llu\n",
Fred Isaman566052c2011-01-06 11:36:20 +00001247 __func__, lseg, lseg->pls_range.iomode,
1248 lseg->pls_range.offset, lseg->pls_range.length,
1249 lp, lp->pls_range.iomode, lp->pls_range.offset,
1250 lp->pls_range.length);
Benny Halevyfb3296e2011-05-22 19:47:26 +03001251 goto out;
Andy Adamson974cec82010-10-20 00:18:02 -04001252 }
Benny Halevyfb3296e2011-05-22 19:47:26 +03001253 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1254 dprintk("%s: inserted lseg %p "
1255 "iomode %d offset %llu length %llu at tail\n",
1256 __func__, lseg, lseg->pls_range.iomode,
1257 lseg->pls_range.offset, lseg->pls_range.length);
1258out:
Trond Myklebust70c3bd22012-09-18 20:51:13 -04001259 pnfs_get_layout_hdr(lo);
Andy Adamson974cec82010-10-20 00:18:02 -04001260
1261 dprintk("%s:Return\n", __func__);
Benny Halevye5e94012010-10-20 00:18:01 -04001262}
Trond Myklebust03772d22015-08-25 08:54:17 -04001263EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1264
1265static void
1266pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1267 struct pnfs_layout_segment *lseg,
1268 struct list_head *free_me)
1269{
1270 struct inode *inode = lo->plh_inode;
1271 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1272
1273 if (ld->add_lseg != NULL)
1274 ld->add_lseg(lo, lseg, free_me);
1275 else
1276 pnfs_generic_layout_insert_lseg(lo, lseg,
1277 pnfs_lseg_range_is_after,
1278 pnfs_lseg_no_merge,
1279 free_me);
1280}
Benny Halevye5e94012010-10-20 00:18:01 -04001281
1282static struct pnfs_layout_hdr *
Peng Tao9fa40752011-07-30 20:52:32 -04001283alloc_init_layout_hdr(struct inode *ino,
1284 struct nfs_open_context *ctx,
1285 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001286{
1287 struct pnfs_layout_hdr *lo;
1288
Benny Halevy636fb9c2011-05-22 19:51:33 +03001289 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
Benny Halevye5e94012010-10-20 00:18:01 -04001290 if (!lo)
1291 return NULL;
Fred Isamancc6e5342011-01-06 11:36:28 +00001292 atomic_set(&lo->plh_refcount, 1);
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001293 INIT_LIST_HEAD(&lo->plh_layouts);
1294 INIT_LIST_HEAD(&lo->plh_segs);
Trond Myklebustfd9a8d72013-02-12 09:48:42 -05001295 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001296 lo->plh_inode = ino;
Trond Myklebust5cc22162013-05-21 09:26:49 -04001297 lo->plh_lc_cred = get_rpccred(ctx->cred);
Benny Halevye5e94012010-10-20 00:18:01 -04001298 return lo;
1299}
1300
1301static struct pnfs_layout_hdr *
Peng Tao9fa40752011-07-30 20:52:32 -04001302pnfs_find_alloc_layout(struct inode *ino,
1303 struct nfs_open_context *ctx,
1304 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001305{
1306 struct nfs_inode *nfsi = NFS_I(ino);
1307 struct pnfs_layout_hdr *new = NULL;
1308
1309 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1310
Trond Myklebust251ec412012-10-02 15:41:05 -07001311 if (nfsi->layout != NULL)
1312 goto out_existing;
Benny Halevye5e94012010-10-20 00:18:01 -04001313 spin_unlock(&ino->i_lock);
Peng Tao9fa40752011-07-30 20:52:32 -04001314 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
Benny Halevye5e94012010-10-20 00:18:01 -04001315 spin_lock(&ino->i_lock);
1316
Trond Myklebust251ec412012-10-02 15:41:05 -07001317 if (likely(nfsi->layout == NULL)) { /* Won the race? */
Benny Halevye5e94012010-10-20 00:18:01 -04001318 nfsi->layout = new;
Trond Myklebust251ec412012-10-02 15:41:05 -07001319 return new;
Yanchuan Nian7175fe92012-10-31 16:05:48 +08001320 } else if (new != NULL)
1321 pnfs_free_layout_hdr(new);
Trond Myklebust251ec412012-10-02 15:41:05 -07001322out_existing:
1323 pnfs_get_layout_hdr(nfsi->layout);
Benny Halevye5e94012010-10-20 00:18:01 -04001324 return nfsi->layout;
1325}
1326
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001327/*
1328 * iomode matching rules:
1329 * iomode lseg match
1330 * ----- ----- -----
1331 * ANY READ true
1332 * ANY RW true
1333 * RW READ false
1334 * RW RW true
1335 * READ READ true
1336 * READ RW true
1337 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001338static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001339pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001340 const struct pnfs_layout_range *range)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001341{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001342 struct pnfs_layout_range range1;
1343
1344 if ((range->iomode == IOMODE_RW &&
1345 ls_range->iomode != IOMODE_RW) ||
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001346 !pnfs_lseg_range_intersecting(ls_range, range))
Benny Halevyfb3296e2011-05-22 19:47:26 +03001347 return 0;
1348
1349 /* range1 covers only the first byte in the range */
1350 range1 = *range;
1351 range1.length = 1;
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001352 return pnfs_lseg_range_contained(ls_range, &range1);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001353}
1354
1355/*
1356 * lookup range in layout
1357 */
Benny Halevye5e94012010-10-20 00:18:01 -04001358static struct pnfs_layout_segment *
Benny Halevyfb3296e2011-05-22 19:47:26 +03001359pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1360 struct pnfs_layout_range *range)
Benny Halevye5e94012010-10-20 00:18:01 -04001361{
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001362 struct pnfs_layout_segment *lseg, *ret = NULL;
1363
1364 dprintk("%s:Begin\n", __func__);
1365
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001366 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
Fred Isaman4541d162011-01-06 11:36:23 +00001367 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
Peng Taoce6ab4f2014-09-06 00:53:24 +08001368 !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001369 pnfs_lseg_range_match(&lseg->pls_range, range)) {
Trond Myklebust9369a432012-09-18 20:57:08 -04001370 ret = pnfs_get_lseg(lseg);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001371 break;
1372 }
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001373 }
1374
1375 dprintk("%s:Return lseg %p ref %d\n",
Fred Isaman4541d162011-01-06 11:36:23 +00001376 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001377 return ret;
Benny Halevye5e94012010-10-20 00:18:01 -04001378}
1379
1380/*
Andy Adamsond23d61c2012-05-23 05:02:37 -04001381 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1382 * to the MDS or over pNFS
1383 *
1384 * The nfs_inode read_io and write_io fields are cumulative counters reset
1385 * when there are no layout segments. Note that in pnfs_update_layout iomode
1386 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1387 * WRITE request.
1388 *
1389 * A return of true means use MDS I/O.
1390 *
1391 * From rfc 5661:
1392 * If a file's size is smaller than the file size threshold, data accesses
1393 * SHOULD be sent to the metadata server. If an I/O request has a length that
1394 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1395 * server. If both file size and I/O size are provided, the client SHOULD
1396 * reach or exceed both thresholds before sending its read or write
1397 * requests to the data server.
1398 */
1399static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1400 struct inode *ino, int iomode)
1401{
1402 struct nfs4_threshold *t = ctx->mdsthreshold;
1403 struct nfs_inode *nfsi = NFS_I(ino);
1404 loff_t fsize = i_size_read(ino);
1405 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1406
1407 if (t == NULL)
1408 return ret;
1409
1410 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1411 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1412
1413 switch (iomode) {
1414 case IOMODE_READ:
1415 if (t->bm & THRESHOLD_RD) {
1416 dprintk("%s fsize %llu\n", __func__, fsize);
1417 size_set = true;
1418 if (fsize < t->rd_sz)
1419 size = true;
1420 }
1421 if (t->bm & THRESHOLD_RD_IO) {
1422 dprintk("%s nfsi->read_io %llu\n", __func__,
1423 nfsi->read_io);
1424 io_set = true;
1425 if (nfsi->read_io < t->rd_io_sz)
1426 io = true;
1427 }
1428 break;
1429 case IOMODE_RW:
1430 if (t->bm & THRESHOLD_WR) {
1431 dprintk("%s fsize %llu\n", __func__, fsize);
1432 size_set = true;
1433 if (fsize < t->wr_sz)
1434 size = true;
1435 }
1436 if (t->bm & THRESHOLD_WR_IO) {
1437 dprintk("%s nfsi->write_io %llu\n", __func__,
1438 nfsi->write_io);
1439 io_set = true;
1440 if (nfsi->write_io < t->wr_io_sz)
1441 io = true;
1442 }
1443 break;
1444 }
1445 if (size_set && io_set) {
1446 if (size && io)
1447 ret = true;
1448 } else if (size || io)
1449 ret = true;
1450
1451 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1452 return ret;
1453}
1454
Peng Taoaa8a45e2014-12-01 08:22:23 +08001455static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1456{
1457 /*
1458 * send layoutcommit as it can hold up layoutreturn due to lseg
1459 * reference
1460 */
1461 pnfs_layoutcommit_inode(lo->plh_inode, false);
1462 return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
Trond Myklebust2e5b29f2015-12-14 16:25:11 -05001463 nfs_wait_bit_killable,
Peng Taoaa8a45e2014-12-01 08:22:23 +08001464 TASK_UNINTERRUPTIBLE);
1465}
1466
Tom Haynesd67ae822014-12-11 17:02:04 -05001467static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1468{
1469 unsigned long *bitlock = &lo->plh_flags;
1470
1471 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1472 smp_mb__after_atomic();
1473 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1474}
1475
Andy Adamsond23d61c2012-05-23 05:02:37 -04001476/*
Benny Halevye5e94012010-10-20 00:18:01 -04001477 * Layout segment is retreived from the server if not cached.
1478 * The appropriate layout segment is referenced and returned to the caller.
1479 */
Andy Adamson7c24d942011-06-13 18:22:38 -04001480struct pnfs_layout_segment *
Benny Halevye5e94012010-10-20 00:18:01 -04001481pnfs_update_layout(struct inode *ino,
1482 struct nfs_open_context *ctx,
Benny Halevyfb3296e2011-05-22 19:47:26 +03001483 loff_t pos,
1484 u64 count,
Trond Myklebusta75b9df2011-05-11 18:00:51 -04001485 enum pnfs_iomode iomode,
1486 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001487{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001488 struct pnfs_layout_range arg = {
1489 .iomode = iomode,
1490 .offset = pos,
1491 .length = count,
1492 };
Benny Halevy707ed5f2011-05-22 19:47:46 +03001493 unsigned pg_offset;
Weston Andros Adamson6382a442011-06-01 16:44:44 -04001494 struct nfs_server *server = NFS_SERVER(ino);
1495 struct nfs_client *clp = server->nfs_client;
Benny Halevye5e94012010-10-20 00:18:01 -04001496 struct pnfs_layout_hdr *lo;
1497 struct pnfs_layout_segment *lseg = NULL;
Weston Andros Adamson30005122013-02-28 20:30:10 -05001498 bool first;
Benny Halevye5e94012010-10-20 00:18:01 -04001499
Jeff Layton9a4bf312015-12-10 10:41:58 -05001500 if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001501 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001502 PNFS_UPDATE_LAYOUT_NO_PNFS);
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001503 goto out;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001504 }
Andy Adamsond23d61c2012-05-23 05:02:37 -04001505
Jeff Layton9a4bf312015-12-10 10:41:58 -05001506 if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001507 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001508 PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
Trond Myklebust4ae935602015-08-31 01:25:11 -07001509 goto out;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001510 }
Trond Myklebust4ae935602015-08-31 01:25:11 -07001511
Jeff Layton9a4bf312015-12-10 10:41:58 -05001512 if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001513 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001514 PNFS_UPDATE_LAYOUT_MDSTHRESH);
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001515 goto out;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001516 }
Andy Adamsond23d61c2012-05-23 05:02:37 -04001517
Peng Tao9bf87482014-08-22 17:37:41 +08001518lookup_again:
1519 first = false;
Benny Halevye5e94012010-10-20 00:18:01 -04001520 spin_lock(&ino->i_lock);
Peng Tao9fa40752011-07-30 20:52:32 -04001521 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001522 if (lo == NULL) {
1523 spin_unlock(&ino->i_lock);
Trond Myklebustf4848302015-12-26 18:03:07 -05001524 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001525 PNFS_UPDATE_LAYOUT_NOMEM);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001526 goto out;
1527 }
Benny Halevye5e94012010-10-20 00:18:01 -04001528
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001529 /* Do we even need to bother with this? */
Trond Myklebusta59c30a2012-03-01 11:17:47 -05001530 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001531 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001532 PNFS_UPDATE_LAYOUT_BULK_RECALL);
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001533 dprintk("%s matches recall, use MDS\n", __func__);
Benny Halevye5e94012010-10-20 00:18:01 -04001534 goto out_unlock;
1535 }
1536
1537 /* if LAYOUTGET already failed once we don't try again */
Trond Myklebust2e5b29f2015-12-14 16:25:11 -05001538 if (pnfs_layout_io_test_failed(lo, iomode)) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001539 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001540 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
Benny Halevye5e94012010-10-20 00:18:01 -04001541 goto out_unlock;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001542 }
Benny Halevye5e94012010-10-20 00:18:01 -04001543
Peng Tao9bf87482014-08-22 17:37:41 +08001544 first = list_empty(&lo->plh_segs);
1545 if (first) {
1546 /* The first layoutget for the file. Need to serialize per
1547 * RFC 5661 Errata 3208.
1548 */
1549 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1550 &lo->plh_flags)) {
1551 spin_unlock(&ino->i_lock);
1552 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1553 TASK_UNINTERRUPTIBLE);
1554 pnfs_put_layout_hdr(lo);
1555 goto lookup_again;
1556 }
1557 } else {
1558 /* Check to see if the layout for the given range
1559 * already exists
1560 */
1561 lseg = pnfs_find_lseg(lo, &arg);
Jeff Layton9a4bf312015-12-10 10:41:58 -05001562 if (lseg) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001563 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001564 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
Peng Tao9bf87482014-08-22 17:37:41 +08001565 goto out_unlock;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001566 }
Peng Tao9bf87482014-08-22 17:37:41 +08001567 }
Andy Adamson568e8c42011-03-01 01:34:22 +00001568
Peng Taoaa8a45e2014-12-01 08:22:23 +08001569 /*
1570 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1571 * for LAYOUTRETURN even if first is true.
1572 */
Trond Myklebust8f70f532015-08-04 16:09:44 -04001573 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
Peng Taoaa8a45e2014-12-01 08:22:23 +08001574 spin_unlock(&ino->i_lock);
1575 dprintk("%s wait for layoutreturn\n", __func__);
1576 if (pnfs_prepare_to_retry_layoutget(lo)) {
Tom Haynesd67ae822014-12-11 17:02:04 -05001577 if (first)
1578 pnfs_clear_first_layoutget(lo);
Peng Taoaa8a45e2014-12-01 08:22:23 +08001579 pnfs_put_layout_hdr(lo);
1580 dprintk("%s retrying\n", __func__);
1581 goto lookup_again;
1582 }
Trond Myklebustf4848302015-12-26 18:03:07 -05001583 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001584 PNFS_UPDATE_LAYOUT_RETURN);
Peng Taoaa8a45e2014-12-01 08:22:23 +08001585 goto out_put_layout_hdr;
1586 }
1587
Jeff Layton9a4bf312015-12-10 10:41:58 -05001588 if (pnfs_layoutgets_blocked(lo)) {
Trond Myklebustf4848302015-12-26 18:03:07 -05001589 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001590 PNFS_UPDATE_LAYOUT_BLOCKED);
Fred Isamancf7d63f2011-01-06 11:36:25 +00001591 goto out_unlock;
Jeff Layton9a4bf312015-12-10 10:41:58 -05001592 }
Fred Isamancf7d63f2011-01-06 11:36:25 +00001593 atomic_inc(&lo->plh_outstanding);
Fred Isamanf49f9ba2011-02-03 18:28:52 +00001594 spin_unlock(&ino->i_lock);
Weston Andros Adamson30005122013-02-28 20:30:10 -05001595
Peng Taoabb9a002014-08-22 17:37:40 +08001596 if (list_empty(&lo->plh_layouts)) {
Fred Isaman2130ff62011-01-06 11:36:26 +00001597 /* The lo must be on the clp list if there is any
1598 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1599 */
1600 spin_lock(&clp->cl_lock);
Peng Taoabb9a002014-08-22 17:37:40 +08001601 if (list_empty(&lo->plh_layouts))
1602 list_add_tail(&lo->plh_layouts, &server->layouts);
Fred Isaman2130ff62011-01-06 11:36:26 +00001603 spin_unlock(&clp->cl_lock);
1604 }
Benny Halevye5e94012010-10-20 00:18:01 -04001605
Benny Halevy707ed5f2011-05-22 19:47:46 +03001606 pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1607 if (pg_offset) {
1608 arg.offset -= pg_offset;
1609 arg.length += pg_offset;
1610 }
Andy Adamson7c24d942011-06-13 18:22:38 -04001611 if (arg.length != NFS4_MAX_UINT64)
1612 arg.length = PAGE_CACHE_ALIGN(arg.length);
Benny Halevy707ed5f2011-05-22 19:47:46 +03001613
Benny Halevyfb3296e2011-05-22 19:47:26 +03001614 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
Fred Isamancf7d63f2011-01-06 11:36:25 +00001615 atomic_dec(&lo->plh_outstanding);
Trond Myklebustf4848302015-12-26 18:03:07 -05001616 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
Jeff Layton9a4bf312015-12-10 10:41:58 -05001617 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001618out_put_layout_hdr:
Tom Haynesd67ae822014-12-11 17:02:04 -05001619 if (first)
1620 pnfs_clear_first_layoutget(lo);
Trond Myklebust70c3bd22012-09-18 20:51:13 -04001621 pnfs_put_layout_hdr(lo);
Benny Halevye5e94012010-10-20 00:18:01 -04001622out:
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001623 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1624 "(%s, offset: %llu, length: %llu)\n",
1625 __func__, ino->i_sb->s_id,
1626 (unsigned long long)NFS_FILEID(ino),
Peng Taod600ad12015-12-04 02:57:48 +08001627 IS_ERR_OR_NULL(lseg) ? "not found" : "found",
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001628 iomode==IOMODE_RW ? "read/write" : "read-only",
1629 (unsigned long long)pos,
1630 (unsigned long long)count);
Benny Halevye5e94012010-10-20 00:18:01 -04001631 return lseg;
1632out_unlock:
1633 spin_unlock(&ino->i_lock);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001634 goto out_put_layout_hdr;
Benny Halevye5e94012010-10-20 00:18:01 -04001635}
Andy Adamson7c24d942011-06-13 18:22:38 -04001636EXPORT_SYMBOL_GPL(pnfs_update_layout);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001637
Trond Myklebust540d9862015-08-25 11:16:13 -04001638static bool
1639pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1640{
1641 switch (range->iomode) {
1642 case IOMODE_READ:
1643 case IOMODE_RW:
1644 break;
1645 default:
1646 return false;
1647 }
1648 if (range->offset == NFS4_MAX_UINT64)
1649 return false;
1650 if (range->length == 0)
1651 return false;
1652 if (range->length != NFS4_MAX_UINT64 &&
1653 range->length > NFS4_MAX_UINT64 - range->offset)
1654 return false;
1655 return true;
1656}
1657
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001658struct pnfs_layout_segment *
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001659pnfs_layout_process(struct nfs4_layoutget *lgp)
1660{
1661 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1662 struct nfs4_layoutget_res *res = &lgp->res;
1663 struct pnfs_layout_segment *lseg;
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001664 struct inode *ino = lo->plh_inode;
Trond Myklebust78096cc2014-02-12 10:02:27 -05001665 LIST_HEAD(free_me);
Trond Myklebust540d9862015-08-25 11:16:13 -04001666 int status = -EINVAL;
1667
1668 if (!pnfs_sanity_check_layout_range(&res->range))
1669 goto out;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001670
1671 /* Inject layout blob into I/O device driver */
Trond Myklebusta75b9df2011-05-11 18:00:51 -04001672 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001673 if (!lseg || IS_ERR(lseg)) {
1674 if (!lseg)
1675 status = -ENOMEM;
1676 else
1677 status = PTR_ERR(lseg);
1678 dprintk("%s: Could not allocate layout: error %d\n",
1679 __func__, status);
1680 goto out;
1681 }
1682
Christoph Hellwig1013df612014-08-21 11:09:18 -05001683 init_lseg(lo, lseg);
1684 lseg->pls_range = res->range;
1685
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001686 spin_lock(&ino->i_lock);
Trond Myklebuste1c06f82015-08-04 16:40:08 -04001687 if (pnfs_layoutgets_blocked(lo)) {
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001688 dprintk("%s forget reply due to state\n", __func__);
1689 goto out_forget_reply;
1690 }
Trond Myklebust038d6492012-10-02 16:38:41 -07001691
Christoph Hellwig362f7472014-08-21 11:09:20 -05001692 if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1693 /* existing state ID, make sure the sequence number matches. */
1694 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1695 dprintk("%s forget reply due to sequence\n", __func__);
Jeff Layton4f2e9dc2015-11-25 13:43:14 -05001696 status = -EAGAIN;
Christoph Hellwig362f7472014-08-21 11:09:20 -05001697 goto out_forget_reply;
1698 }
1699 pnfs_set_layout_stateid(lo, &res->stateid, false);
1700 } else {
1701 /*
1702 * We got an entirely new state ID. Mark all segments for the
1703 * inode invalid, and don't bother validating the stateid
1704 * sequence number.
1705 */
1706 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1707
1708 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1709 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1710 }
Trond Myklebust038d6492012-10-02 16:38:41 -07001711
Christoph Hellwig47abade2014-08-21 11:09:22 -05001712 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1713
Trond Myklebust9369a432012-09-18 20:57:08 -04001714 pnfs_get_lseg(lseg);
Trond Myklebust03772d22015-08-25 08:54:17 -04001715 pnfs_layout_insert_lseg(lo, lseg, &free_me);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001716
Peng Tao39761432015-08-21 12:49:44 +08001717 if (res->return_on_close)
Fred Isamanf7e89172011-01-06 11:36:32 +00001718 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
Fred Isamanf7e89172011-01-06 11:36:32 +00001719
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001720 spin_unlock(&ino->i_lock);
Trond Myklebust78096cc2014-02-12 10:02:27 -05001721 pnfs_free_lseg_list(&free_me);
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001722 return lseg;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001723out:
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001724 return ERR_PTR(status);
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001725
1726out_forget_reply:
1727 spin_unlock(&ino->i_lock);
1728 lseg->pls_layout = lo;
1729 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1730 goto out;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001731}
1732
Peng Tao016256d2014-09-06 00:53:23 +08001733static void
Trond Myklebust5c97f5d2016-01-04 11:53:04 -05001734pnfs_set_plh_return_iomode(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode)
1735{
1736 if (lo->plh_return_iomode == iomode)
1737 return;
1738 if (lo->plh_return_iomode != 0)
1739 iomode = IOMODE_ANY;
1740 lo->plh_return_iomode = iomode;
1741}
1742
Trond Myklebust2f215962016-02-15 12:36:04 -05001743/**
1744 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1745 * @lo: pointer to layout header
1746 * @tmp_list: list header to be used with pnfs_free_lseg_list()
1747 * @return_range: describe layout segment ranges to be returned
1748 *
1749 * This function is mainly intended for use by layoutrecall. It attempts
1750 * to free the layout segment immediately, or else to mark it for return
1751 * as soon as its reference count drops to zero.
1752 */
Trond Myklebust10335552016-01-04 11:23:52 -05001753int
Peng Tao016256d2014-09-06 00:53:23 +08001754pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1755 struct list_head *tmp_list,
Trond Myklebust506c0d62016-01-04 13:04:47 -05001756 const struct pnfs_layout_range *return_range)
Peng Tao016256d2014-09-06 00:53:23 +08001757{
1758 struct pnfs_layout_segment *lseg, *next;
Trond Myklebust10335552016-01-04 11:23:52 -05001759 int remaining = 0;
Peng Tao016256d2014-09-06 00:53:23 +08001760
1761 dprintk("%s:Begin lo %p\n", __func__, lo);
1762
1763 if (list_empty(&lo->plh_segs))
Trond Myklebust10335552016-01-04 11:23:52 -05001764 return 0;
Peng Tao016256d2014-09-06 00:53:23 +08001765
Trond Myklebustfc7ff362015-12-28 10:28:59 -05001766 assert_spin_locked(&lo->plh_inode->i_lock);
Peng Tao016256d2014-09-06 00:53:23 +08001767
1768 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1769 if (should_free_lseg(&lseg->pls_range, return_range)) {
1770 dprintk("%s: marking lseg %p iomode %d "
1771 "offset %llu length %llu\n", __func__,
1772 lseg, lseg->pls_range.iomode,
1773 lseg->pls_range.offset,
1774 lseg->pls_range.length);
Trond Myklebust2f215962016-02-15 12:36:04 -05001775 if (mark_lseg_invalid(lseg, tmp_list))
1776 continue;
1777 remaining++;
Peng Tao016256d2014-09-06 00:53:23 +08001778 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
Trond Myklebust5c97f5d2016-01-04 11:53:04 -05001779 pnfs_set_plh_return_iomode(lo, return_range->iomode);
Trond Myklebust2370abd2016-01-27 20:32:50 -05001780 set_bit(NFS_LAYOUT_RETURN_REQUESTED,
Trond Myklebust0bdb8fa6e2015-08-27 19:17:33 -04001781 &lo->plh_flags);
Peng Tao016256d2014-09-06 00:53:23 +08001782 }
Trond Myklebust10335552016-01-04 11:23:52 -05001783 return remaining;
Peng Tao016256d2014-09-06 00:53:23 +08001784}
1785
1786void pnfs_error_mark_layout_for_return(struct inode *inode,
1787 struct pnfs_layout_segment *lseg)
1788{
1789 struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
Peng Tao016256d2014-09-06 00:53:23 +08001790 struct pnfs_layout_range range = {
1791 .iomode = lseg->pls_range.iomode,
1792 .offset = 0,
1793 .length = NFS4_MAX_UINT64,
1794 };
1795 LIST_HEAD(free_me);
Trond Myklebust10335552016-01-04 11:23:52 -05001796 bool return_now = false;
Peng Tao016256d2014-09-06 00:53:23 +08001797
1798 spin_lock(&inode->i_lock);
Trond Myklebust5c97f5d2016-01-04 11:53:04 -05001799 pnfs_set_plh_return_iomode(lo, range.iomode);
Peng Tao016256d2014-09-06 00:53:23 +08001800 /*
1801 * mark all matching lsegs so that we are sure to have no live
1802 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1803 * for how it works.
1804 */
Trond Myklebust10335552016-01-04 11:23:52 -05001805 if (!pnfs_mark_matching_lsegs_return(lo, &free_me, &range)) {
1806 nfs4_stateid stateid;
1807 enum pnfs_iomode iomode = lo->plh_return_iomode;
1808
1809 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1810 return_now = pnfs_prepare_layoutreturn(lo);
1811 spin_unlock(&inode->i_lock);
1812 if (return_now)
1813 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1814 } else {
1815 spin_unlock(&inode->i_lock);
1816 nfs_commit_inode(inode, 0);
1817 }
Peng Tao016256d2014-09-06 00:53:23 +08001818 pnfs_free_lseg_list(&free_me);
1819}
1820EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1821
Trond Myklebustd8007d42011-06-10 13:30:23 -04001822void
1823pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1824{
Peng Tao1fd937b2012-09-25 14:55:57 +08001825 u64 rd_size = req->wb_bytes;
1826
Peng Taocb5d04b2015-01-24 22:14:52 +08001827 if (pgio->pg_lseg == NULL) {
1828 if (pgio->pg_dreq == NULL)
1829 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1830 else
1831 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
Trond Myklebustd8007d42011-06-10 13:30:23 -04001832
Peng Taocb5d04b2015-01-24 22:14:52 +08001833 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1834 req->wb_context,
1835 req_offset(req),
1836 rd_size,
1837 IOMODE_READ,
1838 GFP_KERNEL);
Peng Taod600ad12015-12-04 02:57:48 +08001839 if (IS_ERR(pgio->pg_lseg)) {
1840 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1841 pgio->pg_lseg = NULL;
1842 return;
1843 }
Peng Taocb5d04b2015-01-24 22:14:52 +08001844 }
Trond Myklebuste885de12011-06-10 13:30:23 -04001845 /* If no lseg, fall back to read through mds */
1846 if (pgio->pg_lseg == NULL)
Trond Myklebust1f945352011-07-13 15:59:57 -04001847 nfs_pageio_reset_read_mds(pgio);
Trond Myklebuste885de12011-06-10 13:30:23 -04001848
Trond Myklebustd8007d42011-06-10 13:30:23 -04001849}
1850EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1851
1852void
Peng Tao62965562012-09-25 14:55:57 +08001853pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1854 struct nfs_page *req, u64 wb_size)
Trond Myklebustd8007d42011-06-10 13:30:23 -04001855{
Peng Taod600ad12015-12-04 02:57:48 +08001856 if (pgio->pg_lseg == NULL) {
Peng Taocb5d04b2015-01-24 22:14:52 +08001857 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1858 req->wb_context,
1859 req_offset(req),
1860 wb_size,
1861 IOMODE_RW,
1862 GFP_NOFS);
Peng Taod600ad12015-12-04 02:57:48 +08001863 if (IS_ERR(pgio->pg_lseg)) {
1864 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1865 pgio->pg_lseg = NULL;
1866 return;
1867 }
1868 }
Trond Myklebuste885de12011-06-10 13:30:23 -04001869 /* If no lseg, fall back to write through mds */
1870 if (pgio->pg_lseg == NULL)
Trond Myklebust1f945352011-07-13 15:59:57 -04001871 nfs_pageio_reset_write_mds(pgio);
Trond Myklebustd8007d42011-06-10 13:30:23 -04001872}
1873EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1874
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04001875void
1876pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1877{
1878 if (desc->pg_lseg) {
1879 pnfs_put_lseg(desc->pg_lseg);
1880 desc->pg_lseg = NULL;
1881 }
1882}
1883EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1884
Weston Andros Adamsonb4fdac12014-05-15 11:56:43 -04001885/*
1886 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1887 * of bytes (maximum @req->wb_bytes) that can be coalesced.
1888 */
1889size_t
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001890pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1891 struct nfs_page *prev, struct nfs_page *req)
Fred Isamanbae724e2011-03-01 01:34:15 +00001892{
Weston Andros Adamson0f9c429e2014-05-15 11:56:51 -04001893 unsigned int size;
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001894 u64 seg_end, req_start, seg_left;
Weston Andros Adamson0f9c429e2014-05-15 11:56:51 -04001895
1896 size = nfs_generic_pg_test(pgio, prev, req);
Weston Andros Adamson0f9c429e2014-05-15 11:56:51 -04001897 if (!size)
1898 return 0;
Benny Halevy89a58e32011-05-25 20:54:40 +03001899
Trond Myklebust19982ba2011-06-10 13:30:23 -04001900 /*
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001901 * 'size' contains the number of bytes left in the current page (up
1902 * to the original size asked for in @req->wb_bytes).
1903 *
1904 * Calculate how many bytes are left in the layout segment
1905 * and if there are less bytes than 'size', return that instead.
Trond Myklebust19982ba2011-06-10 13:30:23 -04001906 *
1907 * Please also note that 'end_offset' is actually the offset of the
1908 * first byte that lies outside the pnfs_layout_range. FIXME?
1909 *
1910 */
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001911 if (pgio->pg_lseg) {
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001912 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1913 pgio->pg_lseg->pls_range.length);
1914 req_start = req_offset(req);
Weston Andros Adamson7c137892015-01-30 11:01:02 -05001915 WARN_ON_ONCE(req_start >= seg_end);
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001916 /* start of request is past the last byte of this segment */
Weston Andros Adamson7c137892015-01-30 11:01:02 -05001917 if (req_start >= seg_end) {
1918 /* reference the new lseg */
1919 if (pgio->pg_ops->pg_cleanup)
1920 pgio->pg_ops->pg_cleanup(pgio);
1921 if (pgio->pg_ops->pg_init)
1922 pgio->pg_ops->pg_init(pgio, req);
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001923 return 0;
Weston Andros Adamson7c137892015-01-30 11:01:02 -05001924 }
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001925
1926 /* adjust 'size' iff there are fewer bytes left in the
1927 * segment than what nfs_generic_pg_test returned */
1928 seg_left = seg_end - req_start;
1929 if (seg_left < size)
1930 size = (unsigned int)seg_left;
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001931 }
Weston Andros Adamson0f9c429e2014-05-15 11:56:51 -04001932
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001933 return size;
Fred Isamanbae724e2011-03-01 01:34:15 +00001934}
Benny Halevy89a58e32011-05-25 20:54:40 +03001935EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
Fred Isamanbae724e2011-03-01 01:34:15 +00001936
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001937int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001938{
1939 struct nfs_pageio_descriptor pgio;
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001940
1941 /* Resend all requests through the MDS */
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001942 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1943 hdr->completion_ops);
Trond Myklebustc7070112015-06-17 19:56:22 -04001944 set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001945 return nfs_pageio_resend(&pgio, hdr);
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001946}
Andy Adamsone7dd79af2012-04-27 17:53:46 -04001947EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001948
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001949static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001950{
Fred Isamancd841602012-04-20 14:47:44 -04001951
1952 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1953 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001954 PNFS_LAYOUTRET_ON_ERROR) {
Fred Isamancd841602012-04-20 14:47:44 -04001955 pnfs_return_layout(hdr->inode);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001956 }
Fred Isaman6c75dc02012-04-20 14:47:47 -04001957 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001958 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001959}
1960
Benny Halevyd20581a2011-05-22 19:52:03 +03001961/*
1962 * Called by non rpc-based layout drivers
1963 */
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001964void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
Fred Isaman94ad1c82011-03-01 01:34:14 +00001965{
Kinglong Meef8417b482015-10-16 17:23:29 +08001966 if (likely(!hdr->pnfs_error)) {
Trond Myklebust67af7612015-03-25 20:40:38 -04001967 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1968 hdr->mds_offset + hdr->res.count);
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001969 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
Kinglong Meef8417b482015-10-16 17:23:29 +08001970 }
1971 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1972 if (unlikely(hdr->pnfs_error))
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001973 pnfs_ld_handle_write_error(hdr);
1974 hdr->mds_ops->rpc_release(hdr);
Fred Isaman44b83792011-03-03 15:13:44 +00001975}
Benny Halevyd20581a2011-05-22 19:52:03 +03001976EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
Fred Isaman44b83792011-03-03 15:13:44 +00001977
Trond Myklebustdce81292011-07-13 15:59:19 -04001978static void
1979pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001980 struct nfs_pgio_header *hdr)
Trond Myklebustdce81292011-07-13 15:59:19 -04001981{
Peng Tao48d635f2014-11-10 08:35:35 +08001982 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001983
Fred Isaman6c75dc02012-04-20 14:47:47 -04001984 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001985 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001986 nfs_pageio_reset_write_mds(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001987 mirror->pg_recoalesce = 1;
Fred Isaman6c75dc02012-04-20 14:47:47 -04001988 }
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001989 nfs_pgio_data_destroy(hdr);
Trond Myklebust1ca018d2015-06-17 19:41:51 -04001990 hdr->release(hdr);
Trond Myklebustdce81292011-07-13 15:59:19 -04001991}
1992
1993static enum pnfs_try_status
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001994pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
Trond Myklebustdce81292011-07-13 15:59:19 -04001995 const struct rpc_call_ops *call_ops,
1996 struct pnfs_layout_segment *lseg,
1997 int how)
Andy Adamson0382b742011-03-03 15:13:45 +00001998{
Fred Isamancd841602012-04-20 14:47:44 -04001999 struct inode *inode = hdr->inode;
Andy Adamson0382b742011-03-03 15:13:45 +00002000 enum pnfs_try_status trypnfs;
2001 struct nfs_server *nfss = NFS_SERVER(inode);
2002
Fred Isamancd841602012-04-20 14:47:44 -04002003 hdr->mds_ops = call_ops;
Andy Adamson0382b742011-03-03 15:13:45 +00002004
2005 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002006 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2007 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
Fred Isaman6c75dc02012-04-20 14:47:47 -04002008 if (trypnfs != PNFS_NOT_ATTEMPTED)
Andy Adamson0382b742011-03-03 15:13:45 +00002009 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
Andy Adamson0382b742011-03-03 15:13:45 +00002010 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2011 return trypnfs;
2012}
2013
Trond Myklebustdce81292011-07-13 15:59:19 -04002014static void
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002015pnfs_do_write(struct nfs_pageio_descriptor *desc,
2016 struct nfs_pgio_header *hdr, int how)
Trond Myklebustdce81292011-07-13 15:59:19 -04002017{
Trond Myklebustdce81292011-07-13 15:59:19 -04002018 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2019 struct pnfs_layout_segment *lseg = desc->pg_lseg;
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002020 enum pnfs_try_status trypnfs;
Trond Myklebustdce81292011-07-13 15:59:19 -04002021
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002022 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002023 if (trypnfs == PNFS_NOT_ATTEMPTED)
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002024 pnfs_write_through_mds(desc, hdr);
Trond Myklebustdce81292011-07-13 15:59:19 -04002025}
2026
Fred Isaman6c75dc02012-04-20 14:47:47 -04002027static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2028{
Trond Myklebust9369a432012-09-18 20:57:08 -04002029 pnfs_put_lseg(hdr->lseg);
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04002030 nfs_pgio_header_free(hdr);
Fred Isaman6c75dc02012-04-20 14:47:47 -04002031}
2032
Trond Myklebustdce81292011-07-13 15:59:19 -04002033int
2034pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2035{
Fred Isaman6c75dc02012-04-20 14:47:47 -04002036 struct nfs_pgio_header *hdr;
Trond Myklebustdce81292011-07-13 15:59:19 -04002037 int ret;
2038
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04002039 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2040 if (!hdr) {
Peng Tao2bff2282015-12-05 02:03:17 +08002041 desc->pg_error = -ENOMEM;
2042 return desc->pg_error;
Fred Isaman6c75dc02012-04-20 14:47:47 -04002043 }
Fred Isaman6c75dc02012-04-20 14:47:47 -04002044 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04002045
Trond Myklebust9369a432012-09-18 20:57:08 -04002046 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
Anna Schumakeref2c4882014-05-06 09:12:36 -04002047 ret = nfs_generic_pgio(desc, hdr);
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04002048 if (!ret)
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002049 pnfs_do_write(desc, hdr, desc->pg_ioflags);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04002050
Fred Isaman6c75dc02012-04-20 14:47:47 -04002051 return ret;
Trond Myklebustdce81292011-07-13 15:59:19 -04002052}
2053EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2054
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04002055int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
Trond Myklebust62e4a762011-11-10 14:30:37 -05002056{
2057 struct nfs_pageio_descriptor pgio;
2058
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002059 /* Resend all requests through the MDS */
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04002060 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2061 return nfs_pageio_resend(&pgio, hdr);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002062}
Andy Adamsone7dd79af2012-04-27 17:53:46 -04002063EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002064
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002065static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002066{
Fred Isamancd841602012-04-20 14:47:44 -04002067 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2068 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002069 PNFS_LAYOUTRET_ON_ERROR) {
Fred Isamancd841602012-04-20 14:47:44 -04002070 pnfs_return_layout(hdr->inode);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04002071 }
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002072 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04002073 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
Trond Myklebust62e4a762011-11-10 14:30:37 -05002074}
2075
Andy Adamsonb1f69b72010-10-20 00:18:03 -04002076/*
Benny Halevyd20581a2011-05-22 19:52:03 +03002077 * Called by non rpc-based layout drivers
2078 */
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002079void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
Benny Halevyd20581a2011-05-22 19:52:03 +03002080{
Fred Isamancd841602012-04-20 14:47:44 -04002081 if (likely(!hdr->pnfs_error)) {
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002082 __nfs4_read_done_cb(hdr);
2083 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
Kinglong Meef8417b482015-10-16 17:23:29 +08002084 }
2085 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2086 if (unlikely(hdr->pnfs_error))
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002087 pnfs_ld_handle_read_error(hdr);
2088 hdr->mds_ops->rpc_release(hdr);
Benny Halevyd20581a2011-05-22 19:52:03 +03002089}
2090EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2091
Trond Myklebust493292d2011-07-13 15:58:28 -04002092static void
2093pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002094 struct nfs_pgio_header *hdr)
Trond Myklebust493292d2011-07-13 15:58:28 -04002095{
Peng Tao48d635f2014-11-10 08:35:35 +08002096 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04002097
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002098 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04002099 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002100 nfs_pageio_reset_read_mds(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04002101 mirror->pg_recoalesce = 1;
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002102 }
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002103 nfs_pgio_data_destroy(hdr);
Trond Myklebust1ca018d2015-06-17 19:41:51 -04002104 hdr->release(hdr);
Trond Myklebust493292d2011-07-13 15:58:28 -04002105}
2106
Benny Halevyd20581a2011-05-22 19:52:03 +03002107/*
Andy Adamson64419a92011-03-01 01:34:16 +00002108 * Call the appropriate parallel I/O subsystem read function.
2109 */
Trond Myklebust493292d2011-07-13 15:58:28 -04002110static enum pnfs_try_status
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002111pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
Trond Myklebust493292d2011-07-13 15:58:28 -04002112 const struct rpc_call_ops *call_ops,
2113 struct pnfs_layout_segment *lseg)
Andy Adamson64419a92011-03-01 01:34:16 +00002114{
Fred Isamancd841602012-04-20 14:47:44 -04002115 struct inode *inode = hdr->inode;
Andy Adamson64419a92011-03-01 01:34:16 +00002116 struct nfs_server *nfss = NFS_SERVER(inode);
2117 enum pnfs_try_status trypnfs;
2118
Fred Isamancd841602012-04-20 14:47:44 -04002119 hdr->mds_ops = call_ops;
Andy Adamson64419a92011-03-01 01:34:16 +00002120
2121 dprintk("%s: Reading ino:%lu %u@%llu\n",
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002122 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
Andy Adamson64419a92011-03-01 01:34:16 +00002123
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002124 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002125 if (trypnfs != PNFS_NOT_ATTEMPTED)
Andy Adamson64419a92011-03-01 01:34:16 +00002126 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
Andy Adamson64419a92011-03-01 01:34:16 +00002127 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2128 return trypnfs;
2129}
Andy Adamson863a3c62011-03-23 13:27:54 +00002130
Peng Taoceb11e12014-11-10 08:35:38 +08002131/* Resend all requests through pnfs. */
2132int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2133{
2134 struct nfs_pageio_descriptor pgio;
2135
2136 nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
2137 return nfs_pageio_resend(&pgio, hdr);
2138}
2139EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2140
Trond Myklebust493292d2011-07-13 15:58:28 -04002141static void
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002142pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
Trond Myklebust493292d2011-07-13 15:58:28 -04002143{
Trond Myklebust493292d2011-07-13 15:58:28 -04002144 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2145 struct pnfs_layout_segment *lseg = desc->pg_lseg;
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002146 enum pnfs_try_status trypnfs;
Peng Taoceb11e12014-11-10 08:35:38 +08002147 int err = 0;
Trond Myklebust493292d2011-07-13 15:58:28 -04002148
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002149 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
Peng Taoceb11e12014-11-10 08:35:38 +08002150 if (trypnfs == PNFS_TRY_AGAIN)
2151 err = pnfs_read_resend_pnfs(hdr);
2152 if (trypnfs == PNFS_NOT_ATTEMPTED || err)
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04002153 pnfs_read_through_mds(desc, hdr);
Trond Myklebust493292d2011-07-13 15:58:28 -04002154}
2155
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002156static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2157{
Trond Myklebust9369a432012-09-18 20:57:08 -04002158 pnfs_put_lseg(hdr->lseg);
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04002159 nfs_pgio_header_free(hdr);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002160}
2161
Trond Myklebust493292d2011-07-13 15:58:28 -04002162int
2163pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2164{
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002165 struct nfs_pgio_header *hdr;
Trond Myklebust493292d2011-07-13 15:58:28 -04002166 int ret;
2167
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04002168 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2169 if (!hdr) {
Peng Tao2bff2282015-12-05 02:03:17 +08002170 desc->pg_error = -ENOMEM;
2171 return desc->pg_error;
Trond Myklebust493292d2011-07-13 15:58:28 -04002172 }
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002173 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
Trond Myklebust9369a432012-09-18 20:57:08 -04002174 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
Anna Schumakeref2c4882014-05-06 09:12:36 -04002175 ret = nfs_generic_pgio(desc, hdr);
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04002176 if (!ret)
Weston Andros Adamson7f714722014-05-15 11:56:53 -04002177 pnfs_do_read(desc, hdr);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04002178 return ret;
Trond Myklebust493292d2011-07-13 15:58:28 -04002179}
2180EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2181
Trond Myklebust71244d92014-01-13 13:34:36 -05002182static void pnfs_clear_layoutcommitting(struct inode *inode)
2183{
2184 unsigned long *bitlock = &NFS_I(inode)->flags;
2185
2186 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
Peter Zijlstra4e857c52014-03-17 18:06:10 +01002187 smp_mb__after_atomic();
Trond Myklebust71244d92014-01-13 13:34:36 -05002188 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2189}
2190
Andy Adamson863a3c62011-03-23 13:27:54 +00002191/*
Peng Taoa9bae562011-07-30 20:52:33 -04002192 * There can be multiple RW segments.
Andy Adamson863a3c62011-03-23 13:27:54 +00002193 */
Peng Taoa9bae562011-07-30 20:52:33 -04002194static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
Andy Adamson863a3c62011-03-23 13:27:54 +00002195{
Peng Taoa9bae562011-07-30 20:52:33 -04002196 struct pnfs_layout_segment *lseg;
Andy Adamson863a3c62011-03-23 13:27:54 +00002197
Peng Taoa9bae562011-07-30 20:52:33 -04002198 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2199 if (lseg->pls_range.iomode == IOMODE_RW &&
Trond Myklebusta073dbf2013-03-20 12:34:32 -04002200 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
Peng Taoa9bae562011-07-30 20:52:33 -04002201 list_add(&lseg->pls_lc_list, listp);
2202 }
Andy Adamson863a3c62011-03-23 13:27:54 +00002203}
2204
Trond Myklebusta073dbf2013-03-20 12:34:32 -04002205static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2206{
2207 struct pnfs_layout_segment *lseg, *tmp;
Trond Myklebusta073dbf2013-03-20 12:34:32 -04002208
2209 /* Matched by references in pnfs_set_layoutcommit */
2210 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2211 list_del_init(&lseg->pls_lc_list);
2212 pnfs_put_lseg(lseg);
2213 }
2214
Trond Myklebust71244d92014-01-13 13:34:36 -05002215 pnfs_clear_layoutcommitting(inode);
Trond Myklebusta073dbf2013-03-20 12:34:32 -04002216}
2217
Peng Tao1b0ae062011-09-22 21:50:12 -04002218void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2219{
Trond Myklebustb9e028f2012-09-18 16:41:18 -04002220 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
Peng Tao1b0ae062011-09-22 21:50:12 -04002221}
2222EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2223
Andy Adamson863a3c62011-03-23 13:27:54 +00002224void
Trond Myklebust67af7612015-03-25 20:40:38 -04002225pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2226 loff_t end_pos)
Andy Adamson863a3c62011-03-23 13:27:54 +00002227{
Fred Isamancd841602012-04-20 14:47:44 -04002228 struct nfs_inode *nfsi = NFS_I(inode);
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04002229 bool mark_as_dirty = false;
Andy Adamson863a3c62011-03-23 13:27:54 +00002230
Fred Isamancd841602012-04-20 14:47:44 -04002231 spin_lock(&inode->i_lock);
Andy Adamson863a3c62011-03-23 13:27:54 +00002232 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
Trond Myklebust29559b12015-03-25 20:10:20 -04002233 nfsi->layout->plh_lwb = end_pos;
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04002234 mark_as_dirty = true;
Andy Adamson863a3c62011-03-23 13:27:54 +00002235 dprintk("%s: Set layoutcommit for inode %lu ",
Fred Isamancd841602012-04-20 14:47:44 -04002236 __func__, inode->i_ino);
Trond Myklebust29559b12015-03-25 20:10:20 -04002237 } else if (end_pos > nfsi->layout->plh_lwb)
2238 nfsi->layout->plh_lwb = end_pos;
Trond Myklebust67af7612015-03-25 20:40:38 -04002239 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
Peng Taoa9bae562011-07-30 20:52:33 -04002240 /* references matched in nfs4_layoutcommit_release */
Trond Myklebust67af7612015-03-25 20:40:38 -04002241 pnfs_get_lseg(lseg);
Peng Taoa9bae562011-07-30 20:52:33 -04002242 }
Fred Isamancd841602012-04-20 14:47:44 -04002243 spin_unlock(&inode->i_lock);
Peng Taoacff58802011-07-30 20:52:31 -04002244 dprintk("%s: lseg %p end_pos %llu\n",
Trond Myklebust67af7612015-03-25 20:40:38 -04002245 __func__, lseg, nfsi->layout->plh_lwb);
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04002246
2247 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2248 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2249 if (mark_as_dirty)
Fred Isamancd841602012-04-20 14:47:44 -04002250 mark_inode_dirty_sync(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00002251}
2252EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2253
Andy Adamsondb29c082011-07-30 20:52:38 -04002254void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2255{
2256 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2257
2258 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2259 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
Trond Myklebusta073dbf2013-03-20 12:34:32 -04002260 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
Andy Adamsondb29c082011-07-30 20:52:38 -04002261}
2262
Andy Adamsonde4b15c2011-03-12 02:58:09 -05002263/*
2264 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2265 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2266 * data to disk to allow the server to recover the data if it crashes.
2267 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2268 * is off, and a COMMIT is sent to a data server, or
2269 * if WRITEs to a data server return NFS_DATA_SYNC.
2270 */
Andy Adamson863a3c62011-03-23 13:27:54 +00002271int
Andy Adamsonef311532011-03-12 02:58:10 -05002272pnfs_layoutcommit_inode(struct inode *inode, bool sync)
Andy Adamson863a3c62011-03-23 13:27:54 +00002273{
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002274 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
Andy Adamson863a3c62011-03-23 13:27:54 +00002275 struct nfs4_layoutcommit_data *data;
2276 struct nfs_inode *nfsi = NFS_I(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00002277 loff_t end_pos;
Trond Myklebust71244d92014-01-13 13:34:36 -05002278 int status;
2279
2280 if (!pnfs_layoutcommit_outstanding(inode))
2281 return 0;
Andy Adamson863a3c62011-03-23 13:27:54 +00002282
2283 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2284
Trond Myklebust71244d92014-01-13 13:34:36 -05002285 status = -EAGAIN;
2286 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2287 if (!sync)
2288 goto out;
NeilBrown74316202014-07-07 15:16:04 +10002289 status = wait_on_bit_lock_action(&nfsi->flags,
Trond Myklebust71244d92014-01-13 13:34:36 -05002290 NFS_INO_LAYOUTCOMMITTING,
2291 nfs_wait_bit_killable,
2292 TASK_KILLABLE);
2293 if (status)
2294 goto out;
2295 }
Andy Adamsonde4b15c2011-03-12 02:58:09 -05002296
Trond Myklebust71244d92014-01-13 13:34:36 -05002297 status = -ENOMEM;
Andy Adamson863a3c62011-03-23 13:27:54 +00002298 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2299 data = kzalloc(sizeof(*data), GFP_NOFS);
Trond Myklebust71244d92014-01-13 13:34:36 -05002300 if (!data)
2301 goto clear_layoutcommitting;
Andy Adamson863a3c62011-03-23 13:27:54 +00002302
Trond Myklebust71244d92014-01-13 13:34:36 -05002303 status = 0;
2304 spin_lock(&inode->i_lock);
2305 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2306 goto out_unlock;
Peng Tao92407e72011-10-23 20:21:17 -07002307
Peng Taoa9bae562011-07-30 20:52:33 -04002308 INIT_LIST_HEAD(&data->lseg_list);
Peng Taoa9bae562011-07-30 20:52:33 -04002309 pnfs_list_write_lseg(inode, &data->lseg_list);
Andy Adamson863a3c62011-03-23 13:27:54 +00002310
Peng Taoacff58802011-07-30 20:52:31 -04002311 end_pos = nfsi->layout->plh_lwb;
Andy Adamson863a3c62011-03-23 13:27:54 +00002312
Trond Myklebustf597c532012-03-04 18:13:56 -05002313 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
Andy Adamson863a3c62011-03-23 13:27:54 +00002314 spin_unlock(&inode->i_lock);
2315
2316 data->args.inode = inode;
Peng Tao9fa40752011-07-30 20:52:32 -04002317 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
Andy Adamson863a3c62011-03-23 13:27:54 +00002318 nfs_fattr_init(&data->fattr);
2319 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2320 data->res.fattr = &data->fattr;
2321 data->args.lastbytewritten = end_pos - 1;
2322 data->res.server = NFS_SERVER(inode);
2323
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002324 if (ld->prepare_layoutcommit) {
2325 status = ld->prepare_layoutcommit(&data->args);
2326 if (status) {
Jeff Layton3471648a2015-07-10 15:58:42 -04002327 put_rpccred(data->cred);
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002328 spin_lock(&inode->i_lock);
Trond Myklebust29559b12015-03-25 20:10:20 -04002329 set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2330 if (end_pos > nfsi->layout->plh_lwb)
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002331 nfsi->layout->plh_lwb = end_pos;
Jeff Layton3471648a2015-07-10 15:58:42 -04002332 goto out_unlock;
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002333 }
2334 }
2335
2336
Andy Adamson863a3c62011-03-23 13:27:54 +00002337 status = nfs4_proc_layoutcommit(data, sync);
2338out:
Peng Tao92407e72011-10-23 20:21:17 -07002339 if (status)
2340 mark_inode_dirty_sync(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00002341 dprintk("<-- %s status %d\n", __func__, status);
2342 return status;
Trond Myklebust71244d92014-01-13 13:34:36 -05002343out_unlock:
2344 spin_unlock(&inode->i_lock);
Peng Tao92407e72011-10-23 20:21:17 -07002345 kfree(data);
Trond Myklebust71244d92014-01-13 13:34:36 -05002346clear_layoutcommitting:
2347 pnfs_clear_layoutcommitting(inode);
Peng Tao92407e72011-10-23 20:21:17 -07002348 goto out;
Andy Adamson863a3c62011-03-23 13:27:54 +00002349}
Peng Tao72cff442014-08-07 10:12:38 +08002350EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
Andy Adamson82be4172012-05-23 05:02:35 -04002351
Trond Myklebust5bb89b42015-03-25 14:14:42 -04002352int
2353pnfs_generic_sync(struct inode *inode, bool datasync)
2354{
2355 return pnfs_layoutcommit_inode(inode, true);
2356}
2357EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2358
Andy Adamson82be4172012-05-23 05:02:35 -04002359struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2360{
2361 struct nfs4_threshold *thp;
2362
2363 thp = kzalloc(sizeof(*thp), GFP_NOFS);
2364 if (!thp) {
2365 dprintk("%s mdsthreshold allocation failed\n", __func__);
2366 return NULL;
2367 }
2368 return thp;
2369}
Peng Tao87334082015-06-23 19:51:57 +08002370
Peng Tao865a7ec2015-06-25 18:19:32 +08002371#if IS_ENABLED(CONFIG_NFS_V4_2)
Peng Tao87334082015-06-23 19:51:57 +08002372int
Trond Myklebustc8ad8892015-08-05 17:31:58 -04002373pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
Peng Tao87334082015-06-23 19:51:57 +08002374{
2375 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2376 struct nfs_server *server = NFS_SERVER(inode);
Peng Tao1bfe3b22015-06-23 19:52:03 +08002377 struct nfs_inode *nfsi = NFS_I(inode);
Peng Tao87334082015-06-23 19:51:57 +08002378 struct nfs42_layoutstat_data *data;
2379 struct pnfs_layout_hdr *hdr;
2380 int status = 0;
2381
2382 if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2383 goto out;
2384
Trond Myklebust6c5a0d82015-06-27 11:45:46 -04002385 if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2386 goto out;
2387
Peng Tao1bfe3b22015-06-23 19:52:03 +08002388 if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2389 goto out;
2390
Peng Tao87334082015-06-23 19:51:57 +08002391 spin_lock(&inode->i_lock);
2392 if (!NFS_I(inode)->layout) {
2393 spin_unlock(&inode->i_lock);
2394 goto out;
2395 }
2396 hdr = NFS_I(inode)->layout;
2397 pnfs_get_layout_hdr(hdr);
2398 spin_unlock(&inode->i_lock);
2399
Trond Myklebustc8ad8892015-08-05 17:31:58 -04002400 data = kzalloc(sizeof(*data), gfp_flags);
Peng Tao87334082015-06-23 19:51:57 +08002401 if (!data) {
2402 status = -ENOMEM;
2403 goto out_put;
2404 }
2405
2406 data->args.fh = NFS_FH(inode);
2407 data->args.inode = inode;
2408 nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2409 status = ld->prepare_layoutstats(&data->args);
2410 if (status)
2411 goto out_free;
2412
2413 status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2414
2415out:
2416 dprintk("%s returns %d\n", __func__, status);
2417 return status;
2418
2419out_free:
2420 kfree(data);
2421out_put:
2422 pnfs_put_layout_hdr(hdr);
Peng Tao1bfe3b22015-06-23 19:52:03 +08002423 smp_mb__before_atomic();
2424 clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2425 smp_mb__after_atomic();
Peng Tao87334082015-06-23 19:51:57 +08002426 goto out;
2427}
2428EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
Peng Tao865a7ec2015-06-25 18:19:32 +08002429#endif
Trond Myklebustbbf58bf2015-08-24 20:39:18 -04002430
2431unsigned int layoutstats_timer;
2432module_param(layoutstats_timer, uint, 0644);
2433EXPORT_SYMBOL_GPL(layoutstats_timer);