blob: fd9536e494bc202184178449b5316c780a1b8e8f [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"
Ricardo Labiaga85e174b2010-10-20 00:17:58 -040037
38#define NFSDBG_FACILITY NFSDBG_PNFS
Trond Myklebust25c75332012-09-18 17:01:12 -040039#define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
Ricardo Labiaga85e174b2010-10-20 00:17:58 -040040
Fred Isaman02c35fc2010-10-20 00:17:59 -040041/* Locking:
42 *
43 * pnfs_spinlock:
44 * protects pnfs_modules_tbl.
45 */
46static DEFINE_SPINLOCK(pnfs_spinlock);
47
48/*
49 * pnfs_modules_tbl holds all pnfs modules
50 */
51static LIST_HEAD(pnfs_modules_tbl);
52
53/* Return the registered pnfs layout driver module matching given id */
54static struct pnfs_layoutdriver_type *
55find_pnfs_driver_locked(u32 id)
56{
57 struct pnfs_layoutdriver_type *local;
58
59 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
60 if (local->id == id)
61 goto out;
62 local = NULL;
63out:
64 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
65 return local;
66}
67
Ricardo Labiaga85e174b2010-10-20 00:17:58 -040068static struct pnfs_layoutdriver_type *
69find_pnfs_driver(u32 id)
70{
Fred Isaman02c35fc2010-10-20 00:17:59 -040071 struct pnfs_layoutdriver_type *local;
72
73 spin_lock(&pnfs_spinlock);
74 local = find_pnfs_driver_locked(id);
Trond Myklebust0a9c63f2012-06-15 13:02:58 -040075 if (local != NULL && !try_module_get(local->owner)) {
76 dprintk("%s: Could not grab reference on module\n", __func__);
77 local = NULL;
78 }
Fred Isaman02c35fc2010-10-20 00:17:59 -040079 spin_unlock(&pnfs_spinlock);
80 return local;
Ricardo Labiaga85e174b2010-10-20 00:17:58 -040081}
82
83void
84unset_pnfs_layoutdriver(struct nfs_server *nfss)
85{
Benny Halevy738fd0f32011-07-30 20:52:36 -040086 if (nfss->pnfs_curr_ld) {
87 if (nfss->pnfs_curr_ld->clear_layoutdriver)
88 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
Trond Myklebust2a4c8992012-06-14 13:08:38 -040089 /* Decrement the MDS count. Purge the deviceid cache if zero */
90 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
91 nfs4_deviceid_purge_client(nfss->nfs_client);
Fred Isaman02c35fc2010-10-20 00:17:59 -040092 module_put(nfss->pnfs_curr_ld->owner);
Benny Halevy738fd0f32011-07-30 20:52:36 -040093 }
Ricardo Labiaga85e174b2010-10-20 00:17:58 -040094 nfss->pnfs_curr_ld = NULL;
95}
96
97/*
98 * Try to set the server's pnfs module to the pnfs layout type specified by id.
99 * Currently only one pNFS layout driver per filesystem is supported.
100 *
101 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
102 */
103void
Benny Halevy738fd0f32011-07-30 20:52:36 -0400104set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
105 u32 id)
Ricardo Labiaga85e174b2010-10-20 00:17:58 -0400106{
107 struct pnfs_layoutdriver_type *ld_type = NULL;
108
109 if (id == 0)
110 goto out_no_driver;
111 if (!(server->nfs_client->cl_exchange_flags &
112 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
Weston Andros Adamsona0308892012-01-26 13:32:23 -0500113 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
114 __func__, id, server->nfs_client->cl_exchange_flags);
Ricardo Labiaga85e174b2010-10-20 00:17:58 -0400115 goto out_no_driver;
116 }
117 ld_type = find_pnfs_driver(id);
118 if (!ld_type) {
119 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
120 ld_type = find_pnfs_driver(id);
121 if (!ld_type) {
122 dprintk("%s: No pNFS module found for %u.\n",
123 __func__, id);
124 goto out_no_driver;
125 }
126 }
127 server->pnfs_curr_ld = ld_type;
Benny Halevy738fd0f32011-07-30 20:52:36 -0400128 if (ld_type->set_layoutdriver
129 && ld_type->set_layoutdriver(server, mntfh)) {
Weston Andros Adamsona0308892012-01-26 13:32:23 -0500130 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
131 "driver %u.\n", __func__, id);
Benny Halevy738fd0f32011-07-30 20:52:36 -0400132 module_put(ld_type->owner);
133 goto out_no_driver;
134 }
Trond Myklebust2a4c8992012-06-14 13:08:38 -0400135 /* Bump the MDS count */
136 atomic_inc(&server->nfs_client->cl_mds_count);
Christoph Hellwigea8eecd2011-03-01 01:34:21 +0000137
Ricardo Labiaga85e174b2010-10-20 00:17:58 -0400138 dprintk("%s: pNFS module for %u set\n", __func__, id);
139 return;
140
141out_no_driver:
142 dprintk("%s: Using NFSv4 I/O\n", __func__);
143 server->pnfs_curr_ld = NULL;
144}
Fred Isaman02c35fc2010-10-20 00:17:59 -0400145
146int
147pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
148{
149 int status = -EINVAL;
150 struct pnfs_layoutdriver_type *tmp;
151
152 if (ld_type->id == 0) {
Weston Andros Adamsona0308892012-01-26 13:32:23 -0500153 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
Fred Isaman02c35fc2010-10-20 00:17:59 -0400154 return status;
155 }
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400156 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
Weston Andros Adamsona0308892012-01-26 13:32:23 -0500157 printk(KERN_ERR "NFS: %s Layout driver must provide "
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400158 "alloc_lseg and free_lseg.\n", __func__);
159 return status;
160 }
Fred Isaman02c35fc2010-10-20 00:17:59 -0400161
162 spin_lock(&pnfs_spinlock);
163 tmp = find_pnfs_driver_locked(ld_type->id);
164 if (!tmp) {
165 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
166 status = 0;
167 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
168 ld_type->name);
169 } else {
Weston Andros Adamsona0308892012-01-26 13:32:23 -0500170 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
Fred Isaman02c35fc2010-10-20 00:17:59 -0400171 __func__, ld_type->id);
172 }
173 spin_unlock(&pnfs_spinlock);
174
175 return status;
176}
177EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
178
179void
180pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
181{
182 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
183 spin_lock(&pnfs_spinlock);
184 list_del(&ld_type->pnfs_tblid);
185 spin_unlock(&pnfs_spinlock);
186}
187EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
Benny Halevye5e94012010-10-20 00:18:01 -0400188
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400189/*
190 * pNFS client layout cache
191 */
192
Fred Isamancc6e5342011-01-06 11:36:28 +0000193/* Need to hold i_lock if caller does not already hold reference */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000194void
Trond Myklebust70c3bd22012-09-18 20:51:13 -0400195pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
Benny Halevye5e94012010-10-20 00:18:01 -0400196{
Fred Isamancc6e5342011-01-06 11:36:28 +0000197 atomic_inc(&lo->plh_refcount);
198}
199
Benny Halevy636fb9c2011-05-22 19:51:33 +0300200static struct pnfs_layout_hdr *
201pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
202{
203 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
Trond Myklebust57934272012-09-20 20:37:23 -0400204 return ld->alloc_layout_hdr(ino, gfp_flags);
Benny Halevy636fb9c2011-05-22 19:51:33 +0300205}
206
207static void
208pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
209{
Trond Myklebust9c626382012-09-20 17:31:43 -0400210 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
211 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
212
213 if (!list_empty(&lo->plh_layouts)) {
214 struct nfs_client *clp = server->nfs_client;
215
216 spin_lock(&clp->cl_lock);
217 list_del_init(&lo->plh_layouts);
218 spin_unlock(&clp->cl_lock);
219 }
Peng Tao9fa40752011-07-30 20:52:32 -0400220 put_rpccred(lo->plh_lc_cred);
Trond Myklebust57934272012-09-20 20:37:23 -0400221 return ld->free_layout_hdr(lo);
Benny Halevy636fb9c2011-05-22 19:51:33 +0300222}
223
Fred Isamancc6e5342011-01-06 11:36:28 +0000224static void
Trond Myklebust6622c3e2012-09-20 17:23:11 -0400225pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
Fred Isamancc6e5342011-01-06 11:36:28 +0000226{
Trond Myklebustbb346f62012-09-20 15:52:13 -0400227 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
Fred Isamancc6e5342011-01-06 11:36:28 +0000228 dprintk("%s: freeing layout cache %p\n", __func__, lo);
Trond Myklebustbb346f62012-09-20 15:52:13 -0400229 nfsi->layout = NULL;
230 /* Reset MDS Threshold I/O counters */
231 nfsi->write_io = 0;
232 nfsi->read_io = 0;
Benny Halevye5e94012010-10-20 00:18:01 -0400233}
234
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400235void
Trond Myklebust70c3bd22012-09-18 20:51:13 -0400236pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
Andy Adamson974cec82010-10-20 00:18:02 -0400237{
Fred Isamancc6e5342011-01-06 11:36:28 +0000238 struct inode *inode = lo->plh_inode;
239
240 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
Trond Myklebust6622c3e2012-09-20 17:23:11 -0400241 pnfs_detach_layout_hdr(lo);
Fred Isamancc6e5342011-01-06 11:36:28 +0000242 spin_unlock(&inode->i_lock);
Trond Myklebust6622c3e2012-09-20 17:23:11 -0400243 pnfs_free_layout_hdr(lo);
Fred Isamancc6e5342011-01-06 11:36:28 +0000244 }
Andy Adamson974cec82010-10-20 00:18:02 -0400245}
246
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400247static int
248pnfs_iomode_to_fail_bit(u32 iomode)
249{
250 return iomode == IOMODE_RW ?
251 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
252}
253
254static void
Trond Myklebust3e621212012-09-24 13:07:16 -0400255pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400256{
Trond Myklebust25c75332012-09-18 17:01:12 -0400257 lo->plh_retry_timestamp = jiffies;
Yanchuan Nian39e88fc2013-01-04 20:19:49 +0800258 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
Trond Myklebust3e621212012-09-24 13:07:16 -0400259 atomic_inc(&lo->plh_refcount);
260}
261
262static void
263pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
264{
265 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
266 atomic_dec(&lo->plh_refcount);
267}
268
269static void
270pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
271{
272 struct inode *inode = lo->plh_inode;
Trond Myklebust115ce572012-09-20 21:19:43 -0400273 struct pnfs_layout_range range = {
274 .iomode = iomode,
275 .offset = 0,
276 .length = NFS4_MAX_UINT64,
277 };
278 LIST_HEAD(head);
Trond Myklebust3e621212012-09-24 13:07:16 -0400279
280 spin_lock(&inode->i_lock);
281 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
Trond Myklebust115ce572012-09-20 21:19:43 -0400282 pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
Trond Myklebust3e621212012-09-24 13:07:16 -0400283 spin_unlock(&inode->i_lock);
Trond Myklebust115ce572012-09-20 21:19:43 -0400284 pnfs_free_lseg_list(&head);
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400285 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
286 iomode == IOMODE_RW ? "RW" : "READ");
287}
288
289static bool
290pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
291{
Trond Myklebust25c75332012-09-18 17:01:12 -0400292 unsigned long start, end;
Trond Myklebust3e621212012-09-24 13:07:16 -0400293 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
294
295 if (test_bit(fail_bit, &lo->plh_flags) == 0)
Trond Myklebust25c75332012-09-18 17:01:12 -0400296 return false;
297 end = jiffies;
298 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
299 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
300 /* It is time to retry the failed layoutgets */
Trond Myklebust3e621212012-09-24 13:07:16 -0400301 pnfs_layout_clear_fail_bit(lo, fail_bit);
Trond Myklebust25c75332012-09-18 17:01:12 -0400302 return false;
303 }
304 return true;
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400305}
306
Andy Adamson974cec82010-10-20 00:18:02 -0400307static void
308init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
309{
Fred Isaman566052c2011-01-06 11:36:20 +0000310 INIT_LIST_HEAD(&lseg->pls_list);
Peng Taoa9bae562011-07-30 20:52:33 -0400311 INIT_LIST_HEAD(&lseg->pls_lc_list);
Fred Isaman4541d162011-01-06 11:36:23 +0000312 atomic_set(&lseg->pls_refcount, 1);
313 smp_mb();
314 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
Fred Isaman566052c2011-01-06 11:36:20 +0000315 lseg->pls_layout = lo;
Andy Adamson974cec82010-10-20 00:18:02 -0400316}
317
Trond Myklebust905ca192012-09-20 20:46:49 -0400318static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
Andy Adamson974cec82010-10-20 00:18:02 -0400319{
Fred Isamanb7edfaa2011-01-06 11:36:21 +0000320 struct inode *ino = lseg->pls_layout->plh_inode;
Andy Adamson974cec82010-10-20 00:18:02 -0400321
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400322 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400323}
324
Fred Isamand684d2a2011-03-01 01:34:13 +0000325static void
Trond Myklebust57036a32012-09-20 16:33:30 -0400326pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
327 struct pnfs_layout_segment *lseg)
Andy Adamson974cec82010-10-20 00:18:02 -0400328{
Trond Myklebust57036a32012-09-20 16:33:30 -0400329 struct inode *inode = lo->plh_inode;
Fred Isamand684d2a2011-03-01 01:34:13 +0000330
Benny Halevyd20581a2011-05-22 19:52:03 +0300331 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
Fred Isamand684d2a2011-03-01 01:34:13 +0000332 list_del_init(&lseg->pls_list);
Trond Myklebust8f0d27d2012-09-20 20:57:11 -0400333 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
334 atomic_dec(&lo->plh_refcount);
Trond Myklebust173f77e2012-09-21 15:49:42 -0400335 if (list_empty(&lo->plh_segs))
336 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
Fred Isamand684d2a2011-03-01 01:34:13 +0000337 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
338}
339
Fred Isamanbae724e2011-03-01 01:34:15 +0000340void
Trond Myklebust9369a432012-09-18 20:57:08 -0400341pnfs_put_lseg(struct pnfs_layout_segment *lseg)
Fred Isamand684d2a2011-03-01 01:34:13 +0000342{
Trond Myklebust57036a32012-09-20 16:33:30 -0400343 struct pnfs_layout_hdr *lo;
Fred Isamand684d2a2011-03-01 01:34:13 +0000344 struct inode *inode;
345
346 if (!lseg)
347 return;
348
Fred Isaman4541d162011-01-06 11:36:23 +0000349 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
350 atomic_read(&lseg->pls_refcount),
351 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
Trond Myklebust57036a32012-09-20 16:33:30 -0400352 lo = lseg->pls_layout;
353 inode = lo->plh_inode;
Fred Isamand684d2a2011-03-01 01:34:13 +0000354 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
Trond Myklebust8f0d27d2012-09-20 20:57:11 -0400355 pnfs_get_layout_hdr(lo);
Trond Myklebust57036a32012-09-20 16:33:30 -0400356 pnfs_layout_remove_lseg(lo, lseg);
Fred Isamand684d2a2011-03-01 01:34:13 +0000357 spin_unlock(&inode->i_lock);
Trond Myklebust905ca192012-09-20 20:46:49 -0400358 pnfs_free_lseg(lseg);
Trond Myklebust8f0d27d2012-09-20 20:57:11 -0400359 pnfs_put_layout_hdr(lo);
Fred Isaman4541d162011-01-06 11:36:23 +0000360 }
Andy Adamson974cec82010-10-20 00:18:02 -0400361}
Trond Myklebust9369a432012-09-18 20:57:08 -0400362EXPORT_SYMBOL_GPL(pnfs_put_lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400363
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400364static u64
Benny Halevyfb3296e2011-05-22 19:47:26 +0300365end_offset(u64 start, u64 len)
Fred Isaman4541d162011-01-06 11:36:23 +0000366{
Benny Halevyfb3296e2011-05-22 19:47:26 +0300367 u64 end;
368
369 end = start + len;
370 return end >= start ? end : NFS4_MAX_UINT64;
371}
372
Benny Halevyfb3296e2011-05-22 19:47:26 +0300373/*
374 * is l2 fully contained in l1?
375 * start1 end1
376 * [----------------------------------)
377 * start2 end2
378 * [----------------)
379 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400380static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400381pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400382 const struct pnfs_layout_range *l2)
Benny Halevyfb3296e2011-05-22 19:47:26 +0300383{
384 u64 start1 = l1->offset;
385 u64 end1 = end_offset(start1, l1->length);
386 u64 start2 = l2->offset;
387 u64 end2 = end_offset(start2, l2->length);
388
389 return (start1 <= start2) && (end1 >= end2);
390}
391
392/*
393 * is l1 and l2 intersecting?
394 * start1 end1
395 * [----------------------------------)
396 * start2 end2
397 * [----------------)
398 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400399static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400400pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400401 const struct pnfs_layout_range *l2)
Benny Halevyfb3296e2011-05-22 19:47:26 +0300402{
403 u64 start1 = l1->offset;
404 u64 end1 = end_offset(start1, l1->length);
405 u64 start2 = l2->offset;
406 u64 end2 = end_offset(start2, l2->length);
407
408 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
409 (end2 == NFS4_MAX_UINT64 || end2 > start1);
410}
411
Fred Isaman4541d162011-01-06 11:36:23 +0000412static bool
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400413should_free_lseg(const struct pnfs_layout_range *lseg_range,
414 const struct pnfs_layout_range *recall_range)
Fred Isaman4541d162011-01-06 11:36:23 +0000415{
Benny Halevy778b5502011-05-22 19:48:02 +0300416 return (recall_range->iomode == IOMODE_ANY ||
417 lseg_range->iomode == recall_range->iomode) &&
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400418 pnfs_lseg_range_intersecting(lseg_range, recall_range);
Fred Isaman4541d162011-01-06 11:36:23 +0000419}
420
Trond Myklebust24956802013-03-20 13:03:00 -0400421static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
422 struct list_head *tmp_list)
423{
424 if (!atomic_dec_and_test(&lseg->pls_refcount))
425 return false;
426 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
427 list_add(&lseg->pls_list, tmp_list);
428 return true;
429}
430
Fred Isaman4541d162011-01-06 11:36:23 +0000431/* Returns 1 if lseg is removed from list, 0 otherwise */
432static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
433 struct list_head *tmp_list)
434{
435 int rv = 0;
436
437 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
438 /* Remove the reference keeping the lseg in the
439 * list. It will now be removed when all
440 * outstanding io is finished.
441 */
Fred Isamand684d2a2011-03-01 01:34:13 +0000442 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
443 atomic_read(&lseg->pls_refcount));
Trond Myklebust24956802013-03-20 13:03:00 -0400444 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
Fred Isamand684d2a2011-03-01 01:34:13 +0000445 rv = 1;
Fred Isaman4541d162011-01-06 11:36:23 +0000446 }
447 return rv;
448}
449
450/* Returns count of number of matching invalid lsegs remaining in list
451 * after call.
452 */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000453int
Trond Myklebust49a85062012-09-18 20:43:31 -0400454pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
Fred Isaman4541d162011-01-06 11:36:23 +0000455 struct list_head *tmp_list,
Benny Halevy778b5502011-05-22 19:48:02 +0300456 struct pnfs_layout_range *recall_range)
Andy Adamson974cec82010-10-20 00:18:02 -0400457{
458 struct pnfs_layout_segment *lseg, *next;
Fred Isaman4541d162011-01-06 11:36:23 +0000459 int invalid = 0, removed = 0;
Andy Adamson974cec82010-10-20 00:18:02 -0400460
461 dprintk("%s:Begin lo %p\n", __func__, lo);
462
Trond Myklebust8006bfb2012-09-21 14:48:04 -0400463 if (list_empty(&lo->plh_segs))
Fred Isaman38511722011-02-03 18:28:50 +0000464 return 0;
Fred Isaman4541d162011-01-06 11:36:23 +0000465 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
Benny Halevy778b5502011-05-22 19:48:02 +0300466 if (!recall_range ||
467 should_free_lseg(&lseg->pls_range, recall_range)) {
Fred Isaman4541d162011-01-06 11:36:23 +0000468 dprintk("%s: freeing lseg %p iomode %d "
469 "offset %llu length %llu\n", __func__,
470 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
471 lseg->pls_range.length);
472 invalid++;
473 removed += mark_lseg_invalid(lseg, tmp_list);
474 }
475 dprintk("%s:Return %i\n", __func__, invalid - removed);
476 return invalid - removed;
Andy Adamson974cec82010-10-20 00:18:02 -0400477}
478
Fred Isamanf49f9ba2011-02-03 18:28:52 +0000479/* note free_me must contain lsegs from a single layout_hdr */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000480void
Fred Isaman4541d162011-01-06 11:36:23 +0000481pnfs_free_lseg_list(struct list_head *free_me)
Andy Adamson974cec82010-10-20 00:18:02 -0400482{
Fred Isaman4541d162011-01-06 11:36:23 +0000483 struct pnfs_layout_segment *lseg, *tmp;
Andy Adamson974cec82010-10-20 00:18:02 -0400484
Fred Isamanf49f9ba2011-02-03 18:28:52 +0000485 if (list_empty(free_me))
486 return;
487
Fred Isaman4541d162011-01-06 11:36:23 +0000488 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
Fred Isaman566052c2011-01-06 11:36:20 +0000489 list_del(&lseg->pls_list);
Trond Myklebust905ca192012-09-20 20:46:49 -0400490 pnfs_free_lseg(lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400491 }
492}
493
Benny Halevye5e94012010-10-20 00:18:01 -0400494void
495pnfs_destroy_layout(struct nfs_inode *nfsi)
496{
497 struct pnfs_layout_hdr *lo;
Andy Adamson974cec82010-10-20 00:18:02 -0400498 LIST_HEAD(tmp_list);
Benny Halevye5e94012010-10-20 00:18:01 -0400499
500 spin_lock(&nfsi->vfs_inode.i_lock);
501 lo = nfsi->layout;
502 if (lo) {
Fred Isaman38511722011-02-03 18:28:50 +0000503 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
Trond Myklebust49a85062012-09-18 20:43:31 -0400504 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
Trond Myklebust3e621212012-09-24 13:07:16 -0400505 pnfs_get_layout_hdr(lo);
506 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
507 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
508 spin_unlock(&nfsi->vfs_inode.i_lock);
509 pnfs_free_lseg_list(&tmp_list);
510 pnfs_put_layout_hdr(lo);
511 } else
512 spin_unlock(&nfsi->vfs_inode.i_lock);
Benny Halevye5e94012010-10-20 00:18:01 -0400513}
Andy Adamson041245c2012-04-27 17:53:53 -0400514EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
Benny Halevye5e94012010-10-20 00:18:01 -0400515
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500516static bool
517pnfs_layout_add_bulk_destroy_list(struct inode *inode,
518 struct list_head *layout_list)
519{
520 struct pnfs_layout_hdr *lo;
521 bool ret = false;
522
523 spin_lock(&inode->i_lock);
524 lo = NFS_I(inode)->layout;
525 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
526 pnfs_get_layout_hdr(lo);
527 list_add(&lo->plh_bulk_destroy, layout_list);
528 ret = true;
529 }
530 spin_unlock(&inode->i_lock);
531 return ret;
532}
533
534/* Caller must hold rcu_read_lock and clp->cl_lock */
535static int
536pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
537 struct nfs_server *server,
538 struct list_head *layout_list)
539{
540 struct pnfs_layout_hdr *lo, *next;
541 struct inode *inode;
542
543 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
544 inode = igrab(lo->plh_inode);
545 if (inode == NULL)
546 continue;
547 list_del_init(&lo->plh_layouts);
548 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
549 continue;
550 rcu_read_unlock();
551 spin_unlock(&clp->cl_lock);
552 iput(inode);
553 spin_lock(&clp->cl_lock);
554 rcu_read_lock();
555 return -EAGAIN;
556 }
557 return 0;
558}
559
560static int
561pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
562 bool is_bulk_recall)
563{
564 struct pnfs_layout_hdr *lo;
565 struct inode *inode;
566 struct pnfs_layout_range range = {
567 .iomode = IOMODE_ANY,
568 .offset = 0,
569 .length = NFS4_MAX_UINT64,
570 };
571 LIST_HEAD(lseg_list);
572 int ret = 0;
573
574 while (!list_empty(layout_list)) {
575 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
576 plh_bulk_destroy);
577 dprintk("%s freeing layout for inode %lu\n", __func__,
578 lo->plh_inode->i_ino);
579 inode = lo->plh_inode;
580 spin_lock(&inode->i_lock);
581 list_del_init(&lo->plh_bulk_destroy);
582 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
583 if (is_bulk_recall)
584 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
585 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
586 ret = -EAGAIN;
587 spin_unlock(&inode->i_lock);
588 pnfs_free_lseg_list(&lseg_list);
589 pnfs_put_layout_hdr(lo);
590 iput(inode);
591 }
592 return ret;
593}
594
595int
596pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
597 struct nfs_fsid *fsid,
598 bool is_recall)
599{
600 struct nfs_server *server;
601 LIST_HEAD(layout_list);
602
603 spin_lock(&clp->cl_lock);
604 rcu_read_lock();
605restart:
606 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
607 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
608 continue;
609 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
610 server,
611 &layout_list) != 0)
612 goto restart;
613 }
614 rcu_read_unlock();
615 spin_unlock(&clp->cl_lock);
616
617 if (list_empty(&layout_list))
618 return 0;
619 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
620}
621
622int
623pnfs_destroy_layouts_byclid(struct nfs_client *clp,
624 bool is_recall)
625{
626 struct nfs_server *server;
627 LIST_HEAD(layout_list);
628
629 spin_lock(&clp->cl_lock);
630 rcu_read_lock();
631restart:
632 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
633 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
634 server,
635 &layout_list) != 0)
636 goto restart;
637 }
638 rcu_read_unlock();
639 spin_unlock(&clp->cl_lock);
640
641 if (list_empty(&layout_list))
642 return 0;
643 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
644}
645
Andy Adamson974cec82010-10-20 00:18:02 -0400646/*
647 * Called by the state manger to remove all layouts established under an
648 * expired lease.
649 */
650void
651pnfs_destroy_all_layouts(struct nfs_client *clp)
652{
Andy Adamsonc47abcf2011-06-15 17:52:40 -0400653 nfs4_deviceid_mark_client_invalid(clp);
654 nfs4_deviceid_purge_client(clp);
655
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500656 pnfs_destroy_layouts_byclid(clp, false);
Andy Adamson974cec82010-10-20 00:18:02 -0400657}
658
Trond Myklebust5a655032012-10-02 16:47:14 -0700659/*
660 * Compare 2 layout stateid sequence ids, to see which is newer,
661 * taking into account wraparound issues.
662 */
663static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
664{
Trond Myklebust2c64c572014-02-12 09:36:59 -0500665 return (s32)(s1 - s2) > 0;
Trond Myklebust5a655032012-10-02 16:47:14 -0700666}
667
Trond Myklebust78096cc2014-02-12 10:02:27 -0500668static void
669pnfs_verify_layout_stateid(struct pnfs_layout_hdr *lo,
670 const nfs4_stateid *new,
671 struct list_head *free_me_list)
672{
673 if (nfs4_stateid_match_other(&lo->plh_stateid, new))
674 return;
675 /* Layout is new! Kill existing layout segments */
676 pnfs_mark_matching_lsegs_invalid(lo, free_me_list, NULL);
677}
678
Fred Isamanfd6002e2011-01-06 11:36:22 +0000679/* update lo->plh_stateid with new if is more recent */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000680void
681pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
682 bool update_barrier)
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400683{
Trond Myklebust22aaf712012-10-04 16:32:22 -0700684 u32 oldseq, newseq, new_barrier;
685 int empty = list_empty(&lo->plh_segs);
Andy Adamson974cec82010-10-20 00:18:02 -0400686
Trond Myklebust2d2f24a2012-03-04 18:13:57 -0500687 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
688 newseq = be32_to_cpu(new->seqid);
Trond Myklebust22aaf712012-10-04 16:32:22 -0700689 if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
Trond Myklebustf597c532012-03-04 18:13:56 -0500690 nfs4_stateid_copy(&lo->plh_stateid, new);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000691 if (update_barrier) {
Trond Myklebust22aaf712012-10-04 16:32:22 -0700692 new_barrier = be32_to_cpu(new->seqid);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000693 } else {
694 /* Because of wraparound, we want to keep the barrier
Trond Myklebust22aaf712012-10-04 16:32:22 -0700695 * "close" to the current seqids.
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000696 */
Trond Myklebust22aaf712012-10-04 16:32:22 -0700697 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000698 }
Trond Myklebust22aaf712012-10-04 16:32:22 -0700699 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
700 lo->plh_barrier = new_barrier;
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000701 }
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400702}
703
Trond Myklebust19c54ab2012-10-05 16:56:58 -0700704static bool
705pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
706 const nfs4_stateid *stateid)
707{
708 u32 seqid = be32_to_cpu(stateid->seqid);
709
710 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
711}
712
Fred Isamancf7d63f2011-01-06 11:36:25 +0000713/* lget is set to 1 if called from inside send_layoutget call chain */
714static bool
Trond Myklebust19c54ab2012-10-05 16:56:58 -0700715pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
Fred Isamancf7d63f2011-01-06 11:36:25 +0000716{
Fred Isamanf7e89172011-01-06 11:36:32 +0000717 return lo->plh_block_lgets ||
718 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000719 (list_empty(&lo->plh_segs) &&
Fred Isamancf7d63f2011-01-06 11:36:25 +0000720 (atomic_read(&lo->plh_outstanding) > lget));
721}
722
Fred Isamanfd6002e2011-01-06 11:36:22 +0000723int
724pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
725 struct nfs4_state *open_state)
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400726{
Fred Isamanfd6002e2011-01-06 11:36:22 +0000727 int status = 0;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400728
729 dprintk("--> %s\n", __func__);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000730 spin_lock(&lo->plh_inode->i_lock);
Trond Myklebust19c54ab2012-10-05 16:56:58 -0700731 if (pnfs_layoutgets_blocked(lo, 1)) {
Fred Isamancf7d63f2011-01-06 11:36:25 +0000732 status = -EAGAIN;
Trond Myklebust5d422302013-03-14 16:57:48 -0400733 } else if (!nfs4_valid_open_stateid(open_state)) {
734 status = -EBADF;
Fred Isamancf7d63f2011-01-06 11:36:25 +0000735 } else if (list_empty(&lo->plh_segs)) {
Fred Isamanfd6002e2011-01-06 11:36:22 +0000736 int seq;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400737
Fred Isamanfd6002e2011-01-06 11:36:22 +0000738 do {
739 seq = read_seqbegin(&open_state->seqlock);
Trond Myklebustf597c532012-03-04 18:13:56 -0500740 nfs4_stateid_copy(dst, &open_state->stateid);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000741 } while (read_seqretry(&open_state->seqlock, seq));
742 } else
Trond Myklebustf597c532012-03-04 18:13:56 -0500743 nfs4_stateid_copy(dst, &lo->plh_stateid);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000744 spin_unlock(&lo->plh_inode->i_lock);
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400745 dprintk("<-- %s\n", __func__);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000746 return status;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400747}
748
749/*
750* Get layout from server.
751* for now, assume that whole file layouts are requested.
752* arg->offset: 0
753* arg->length: all ones
754*/
Benny Halevye5e94012010-10-20 00:18:01 -0400755static struct pnfs_layout_segment *
756send_layoutget(struct pnfs_layout_hdr *lo,
757 struct nfs_open_context *ctx,
Benny Halevyfb3296e2011-05-22 19:47:26 +0300758 struct pnfs_layout_range *range,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400759 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -0400760{
Fred Isamanb7edfaa2011-01-06 11:36:21 +0000761 struct inode *ino = lo->plh_inode;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400762 struct nfs_server *server = NFS_SERVER(ino);
763 struct nfs4_layoutget *lgp;
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -0400764 struct pnfs_layout_segment *lseg;
Benny Halevye5e94012010-10-20 00:18:01 -0400765
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400766 dprintk("--> %s\n", __func__);
767
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400768 lgp = kzalloc(sizeof(*lgp), gfp_flags);
Fred Isamancf7d63f2011-01-06 11:36:25 +0000769 if (lgp == NULL)
Andy Adamson974cec82010-10-20 00:18:02 -0400770 return NULL;
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400771
Benny Halevyfb3296e2011-05-22 19:47:26 +0300772 lgp->args.minlength = PAGE_CACHE_SIZE;
773 if (lgp->args.minlength > range->length)
774 lgp->args.minlength = range->length;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400775 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
Benny Halevyfb3296e2011-05-22 19:47:26 +0300776 lgp->args.range = *range;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400777 lgp->args.type = server->pnfs_curr_ld->id;
778 lgp->args.inode = ino;
779 lgp->args.ctx = get_nfs_open_context(ctx);
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400780 lgp->gfp_flags = gfp_flags;
Trond Myklebust6ab59342013-05-20 10:49:34 -0400781 lgp->cred = lo->plh_lc_cred;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400782
783 /* Synchronously retrieve layout information from server and
784 * store in lseg.
785 */
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -0400786 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
787 if (IS_ERR(lseg)) {
788 switch (PTR_ERR(lseg)) {
789 case -ENOMEM:
790 case -ERESTARTSYS:
791 break;
792 default:
793 /* remember that LAYOUTGET failed and suspend trying */
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400794 pnfs_layout_io_set_failed(lo, range->iomode);
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -0400795 }
796 return NULL;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400797 }
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400798
Andy Adamson974cec82010-10-20 00:18:02 -0400799 return lseg;
800}
801
Trond Myklebust24956802013-03-20 13:03:00 -0400802static void pnfs_clear_layoutcommit(struct inode *inode,
803 struct list_head *head)
804{
805 struct nfs_inode *nfsi = NFS_I(inode);
806 struct pnfs_layout_segment *lseg, *tmp;
807
808 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
809 return;
810 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
811 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
812 continue;
813 pnfs_lseg_dec_and_remove_zero(lseg, head);
814 }
815}
816
Andy Adamson293b3b02012-06-20 15:03:34 -0400817/*
818 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
819 * when the layout segment list is empty.
820 *
821 * Note that a pnfs_layout_hdr can exist with an empty layout segment
822 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
823 * deviceid is marked invalid.
824 */
Benny Halevycbe82602011-05-22 19:52:37 +0300825int
826_pnfs_return_layout(struct inode *ino)
827{
828 struct pnfs_layout_hdr *lo = NULL;
829 struct nfs_inode *nfsi = NFS_I(ino);
830 LIST_HEAD(tmp_list);
831 struct nfs4_layoutreturn *lrp;
832 nfs4_stateid stateid;
Andy Adamson293b3b02012-06-20 15:03:34 -0400833 int status = 0, empty;
Benny Halevycbe82602011-05-22 19:52:37 +0300834
Andy Adamson366d5052012-06-20 15:03:33 -0400835 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
Benny Halevycbe82602011-05-22 19:52:37 +0300836
837 spin_lock(&ino->i_lock);
838 lo = nfsi->layout;
Trond Myklebuste5929f32012-09-21 16:37:02 -0400839 if (!lo) {
Benny Halevycbe82602011-05-22 19:52:37 +0300840 spin_unlock(&ino->i_lock);
Andy Adamson293b3b02012-06-20 15:03:34 -0400841 dprintk("NFS: %s no layout to return\n", __func__);
842 goto out;
Benny Halevycbe82602011-05-22 19:52:37 +0300843 }
844 stateid = nfsi->layout->plh_stateid;
845 /* Reference matched in nfs4_layoutreturn_release */
Trond Myklebust70c3bd22012-09-18 20:51:13 -0400846 pnfs_get_layout_hdr(lo);
Andy Adamson293b3b02012-06-20 15:03:34 -0400847 empty = list_empty(&lo->plh_segs);
Trond Myklebust24956802013-03-20 13:03:00 -0400848 pnfs_clear_layoutcommit(ino, &tmp_list);
Trond Myklebust49a85062012-09-18 20:43:31 -0400849 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
Andy Adamson293b3b02012-06-20 15:03:34 -0400850 /* Don't send a LAYOUTRETURN if list was initially empty */
851 if (empty) {
852 spin_unlock(&ino->i_lock);
Trond Myklebust70c3bd22012-09-18 20:51:13 -0400853 pnfs_put_layout_hdr(lo);
Andy Adamson293b3b02012-06-20 15:03:34 -0400854 dprintk("NFS: %s no layout segments to return\n", __func__);
855 goto out;
856 }
Fred Isamanea0ded72011-06-15 12:31:02 -0400857 lo->plh_block_lgets++;
Benny Halevycbe82602011-05-22 19:52:37 +0300858 spin_unlock(&ino->i_lock);
859 pnfs_free_lseg_list(&tmp_list);
860
Benny Halevycbe82602011-05-22 19:52:37 +0300861 lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
862 if (unlikely(lrp == NULL)) {
863 status = -ENOMEM;
Trond Myklebust65857d52012-09-24 13:49:27 -0400864 spin_lock(&ino->i_lock);
865 lo->plh_block_lgets--;
866 spin_unlock(&ino->i_lock);
Trond Myklebust70c3bd22012-09-18 20:51:13 -0400867 pnfs_put_layout_hdr(lo);
Benny Halevycbe82602011-05-22 19:52:37 +0300868 goto out;
869 }
870
871 lrp->args.stateid = stateid;
872 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
873 lrp->args.inode = ino;
Trond Myklebusta56aaa02011-06-15 11:59:10 -0400874 lrp->args.layout = lo;
Benny Halevycbe82602011-05-22 19:52:37 +0300875 lrp->clp = NFS_SERVER(ino)->nfs_client;
Trond Myklebust95560002013-05-20 10:43:47 -0400876 lrp->cred = lo->plh_lc_cred;
Benny Halevycbe82602011-05-22 19:52:37 +0300877
878 status = nfs4_proc_layoutreturn(lrp);
879out:
880 dprintk("<-- %s status: %d\n", __func__, status);
881 return status;
882}
Andy Adamson0a57cda2012-04-27 17:53:50 -0400883EXPORT_SYMBOL_GPL(_pnfs_return_layout);
Benny Halevycbe82602011-05-22 19:52:37 +0300884
Trond Myklebust24028672013-03-20 13:23:33 -0400885int
886pnfs_commit_and_return_layout(struct inode *inode)
887{
888 struct pnfs_layout_hdr *lo;
889 int ret;
890
891 spin_lock(&inode->i_lock);
892 lo = NFS_I(inode)->layout;
893 if (lo == NULL) {
894 spin_unlock(&inode->i_lock);
895 return 0;
896 }
897 pnfs_get_layout_hdr(lo);
898 /* Block new layoutgets and read/write to ds */
899 lo->plh_block_lgets++;
900 spin_unlock(&inode->i_lock);
901 filemap_fdatawait(inode->i_mapping);
902 ret = pnfs_layoutcommit_inode(inode, true);
903 if (ret == 0)
904 ret = _pnfs_return_layout(inode);
905 spin_lock(&inode->i_lock);
906 lo->plh_block_lgets--;
907 spin_unlock(&inode->i_lock);
908 pnfs_put_layout_hdr(lo);
909 return ret;
910}
911
Fred Isamanf7e89172011-01-06 11:36:32 +0000912bool pnfs_roc(struct inode *ino)
913{
914 struct pnfs_layout_hdr *lo;
915 struct pnfs_layout_segment *lseg, *tmp;
916 LIST_HEAD(tmp_list);
917 bool found = false;
918
919 spin_lock(&ino->i_lock);
920 lo = NFS_I(ino)->layout;
921 if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
922 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
923 goto out_nolayout;
924 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
925 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
926 mark_lseg_invalid(lseg, &tmp_list);
927 found = true;
928 }
929 if (!found)
930 goto out_nolayout;
931 lo->plh_block_lgets++;
Trond Myklebust70c3bd22012-09-18 20:51:13 -0400932 pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
Fred Isamanf7e89172011-01-06 11:36:32 +0000933 spin_unlock(&ino->i_lock);
934 pnfs_free_lseg_list(&tmp_list);
935 return true;
936
937out_nolayout:
938 spin_unlock(&ino->i_lock);
939 return false;
940}
941
942void pnfs_roc_release(struct inode *ino)
943{
944 struct pnfs_layout_hdr *lo;
945
946 spin_lock(&ino->i_lock);
947 lo = NFS_I(ino)->layout;
948 lo->plh_block_lgets--;
Trond Myklebust6622c3e2012-09-20 17:23:11 -0400949 if (atomic_dec_and_test(&lo->plh_refcount)) {
950 pnfs_detach_layout_hdr(lo);
951 spin_unlock(&ino->i_lock);
952 pnfs_free_layout_hdr(lo);
953 } else
954 spin_unlock(&ino->i_lock);
Fred Isamanf7e89172011-01-06 11:36:32 +0000955}
956
957void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
958{
959 struct pnfs_layout_hdr *lo;
960
961 spin_lock(&ino->i_lock);
962 lo = NFS_I(ino)->layout;
Trond Myklebust0f35ad62012-10-04 16:28:17 -0700963 if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
Fred Isamanf7e89172011-01-06 11:36:32 +0000964 lo->plh_barrier = barrier;
965 spin_unlock(&ino->i_lock);
966}
967
Trond Myklebust7fdab062012-09-20 20:15:57 -0400968bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
Fred Isamanf7e89172011-01-06 11:36:32 +0000969{
970 struct nfs_inode *nfsi = NFS_I(ino);
Trond Myklebust7fdab062012-09-20 20:15:57 -0400971 struct pnfs_layout_hdr *lo;
Fred Isamanf7e89172011-01-06 11:36:32 +0000972 struct pnfs_layout_segment *lseg;
Trond Myklebust7fdab062012-09-20 20:15:57 -0400973 u32 current_seqid;
Fred Isamanf7e89172011-01-06 11:36:32 +0000974 bool found = false;
975
976 spin_lock(&ino->i_lock);
977 list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
978 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
Trond Myklebust7fdab062012-09-20 20:15:57 -0400979 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
Fred Isamanf7e89172011-01-06 11:36:32 +0000980 found = true;
Trond Myklebust7fdab062012-09-20 20:15:57 -0400981 goto out;
Fred Isamanf7e89172011-01-06 11:36:32 +0000982 }
Trond Myklebust7fdab062012-09-20 20:15:57 -0400983 lo = nfsi->layout;
984 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
Fred Isamanf7e89172011-01-06 11:36:32 +0000985
Trond Myklebust7fdab062012-09-20 20:15:57 -0400986 /* Since close does not return a layout stateid for use as
987 * a barrier, we choose the worst-case barrier.
988 */
989 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
990out:
Fred Isamanf7e89172011-01-06 11:36:32 +0000991 spin_unlock(&ino->i_lock);
992 return found;
993}
994
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400995/*
996 * Compare two layout segments for sorting into layout cache.
997 * We want to preferentially return RW over RO layouts, so ensure those
998 * are seen first.
999 */
1000static s64
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001001pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001002 const struct pnfs_layout_range *l2)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001003{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001004 s64 d;
1005
1006 /* high offset > low offset */
1007 d = l1->offset - l2->offset;
1008 if (d)
1009 return d;
1010
1011 /* short length > long length */
1012 d = l2->length - l1->length;
1013 if (d)
1014 return d;
1015
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001016 /* read > read/write */
Benny Halevyfb3296e2011-05-22 19:47:26 +03001017 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001018}
1019
Andy Adamson974cec82010-10-20 00:18:02 -04001020static void
Trond Myklebust57036a32012-09-20 16:33:30 -04001021pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
Andy Adamson974cec82010-10-20 00:18:02 -04001022 struct pnfs_layout_segment *lseg)
1023{
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001024 struct pnfs_layout_segment *lp;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001025
Andy Adamson974cec82010-10-20 00:18:02 -04001026 dprintk("%s:Begin\n", __func__);
1027
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001028 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001029 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001030 continue;
Fred Isaman566052c2011-01-06 11:36:20 +00001031 list_add_tail(&lseg->pls_list, &lp->pls_list);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001032 dprintk("%s: inserted lseg %p "
1033 "iomode %d offset %llu length %llu before "
1034 "lp %p iomode %d offset %llu length %llu\n",
Fred Isaman566052c2011-01-06 11:36:20 +00001035 __func__, lseg, lseg->pls_range.iomode,
1036 lseg->pls_range.offset, lseg->pls_range.length,
1037 lp, lp->pls_range.iomode, lp->pls_range.offset,
1038 lp->pls_range.length);
Benny Halevyfb3296e2011-05-22 19:47:26 +03001039 goto out;
Andy Adamson974cec82010-10-20 00:18:02 -04001040 }
Benny Halevyfb3296e2011-05-22 19:47:26 +03001041 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1042 dprintk("%s: inserted lseg %p "
1043 "iomode %d offset %llu length %llu at tail\n",
1044 __func__, lseg, lseg->pls_range.iomode,
1045 lseg->pls_range.offset, lseg->pls_range.length);
1046out:
Trond Myklebust70c3bd22012-09-18 20:51:13 -04001047 pnfs_get_layout_hdr(lo);
Andy Adamson974cec82010-10-20 00:18:02 -04001048
1049 dprintk("%s:Return\n", __func__);
Benny Halevye5e94012010-10-20 00:18:01 -04001050}
1051
1052static struct pnfs_layout_hdr *
Peng Tao9fa40752011-07-30 20:52:32 -04001053alloc_init_layout_hdr(struct inode *ino,
1054 struct nfs_open_context *ctx,
1055 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001056{
1057 struct pnfs_layout_hdr *lo;
1058
Benny Halevy636fb9c2011-05-22 19:51:33 +03001059 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
Benny Halevye5e94012010-10-20 00:18:01 -04001060 if (!lo)
1061 return NULL;
Fred Isamancc6e5342011-01-06 11:36:28 +00001062 atomic_set(&lo->plh_refcount, 1);
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001063 INIT_LIST_HEAD(&lo->plh_layouts);
1064 INIT_LIST_HEAD(&lo->plh_segs);
Trond Myklebustfd9a8d72013-02-12 09:48:42 -05001065 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001066 lo->plh_inode = ino;
Trond Myklebust5cc22162013-05-21 09:26:49 -04001067 lo->plh_lc_cred = get_rpccred(ctx->cred);
Benny Halevye5e94012010-10-20 00:18:01 -04001068 return lo;
1069}
1070
1071static struct pnfs_layout_hdr *
Peng Tao9fa40752011-07-30 20:52:32 -04001072pnfs_find_alloc_layout(struct inode *ino,
1073 struct nfs_open_context *ctx,
1074 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001075{
1076 struct nfs_inode *nfsi = NFS_I(ino);
1077 struct pnfs_layout_hdr *new = NULL;
1078
1079 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1080
Trond Myklebust251ec412012-10-02 15:41:05 -07001081 if (nfsi->layout != NULL)
1082 goto out_existing;
Benny Halevye5e94012010-10-20 00:18:01 -04001083 spin_unlock(&ino->i_lock);
Peng Tao9fa40752011-07-30 20:52:32 -04001084 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
Benny Halevye5e94012010-10-20 00:18:01 -04001085 spin_lock(&ino->i_lock);
1086
Trond Myklebust251ec412012-10-02 15:41:05 -07001087 if (likely(nfsi->layout == NULL)) { /* Won the race? */
Benny Halevye5e94012010-10-20 00:18:01 -04001088 nfsi->layout = new;
Trond Myklebust251ec412012-10-02 15:41:05 -07001089 return new;
Yanchuan Nian7175fe92012-10-31 16:05:48 +08001090 } else if (new != NULL)
1091 pnfs_free_layout_hdr(new);
Trond Myklebust251ec412012-10-02 15:41:05 -07001092out_existing:
1093 pnfs_get_layout_hdr(nfsi->layout);
Benny Halevye5e94012010-10-20 00:18:01 -04001094 return nfsi->layout;
1095}
1096
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001097/*
1098 * iomode matching rules:
1099 * iomode lseg match
1100 * ----- ----- -----
1101 * ANY READ true
1102 * ANY RW true
1103 * RW READ false
1104 * RW RW true
1105 * READ READ true
1106 * READ RW true
1107 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001108static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001109pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001110 const struct pnfs_layout_range *range)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001111{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001112 struct pnfs_layout_range range1;
1113
1114 if ((range->iomode == IOMODE_RW &&
1115 ls_range->iomode != IOMODE_RW) ||
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001116 !pnfs_lseg_range_intersecting(ls_range, range))
Benny Halevyfb3296e2011-05-22 19:47:26 +03001117 return 0;
1118
1119 /* range1 covers only the first byte in the range */
1120 range1 = *range;
1121 range1.length = 1;
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001122 return pnfs_lseg_range_contained(ls_range, &range1);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001123}
1124
1125/*
1126 * lookup range in layout
1127 */
Benny Halevye5e94012010-10-20 00:18:01 -04001128static struct pnfs_layout_segment *
Benny Halevyfb3296e2011-05-22 19:47:26 +03001129pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1130 struct pnfs_layout_range *range)
Benny Halevye5e94012010-10-20 00:18:01 -04001131{
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001132 struct pnfs_layout_segment *lseg, *ret = NULL;
1133
1134 dprintk("%s:Begin\n", __func__);
1135
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001136 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
Fred Isaman4541d162011-01-06 11:36:23 +00001137 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001138 pnfs_lseg_range_match(&lseg->pls_range, range)) {
Trond Myklebust9369a432012-09-18 20:57:08 -04001139 ret = pnfs_get_lseg(lseg);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001140 break;
1141 }
Benny Halevyd771e3a2011-06-14 16:30:16 -04001142 if (lseg->pls_range.offset > range->offset)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001143 break;
1144 }
1145
1146 dprintk("%s:Return lseg %p ref %d\n",
Fred Isaman4541d162011-01-06 11:36:23 +00001147 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001148 return ret;
Benny Halevye5e94012010-10-20 00:18:01 -04001149}
1150
1151/*
Andy Adamsond23d61c2012-05-23 05:02:37 -04001152 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1153 * to the MDS or over pNFS
1154 *
1155 * The nfs_inode read_io and write_io fields are cumulative counters reset
1156 * when there are no layout segments. Note that in pnfs_update_layout iomode
1157 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1158 * WRITE request.
1159 *
1160 * A return of true means use MDS I/O.
1161 *
1162 * From rfc 5661:
1163 * If a file's size is smaller than the file size threshold, data accesses
1164 * SHOULD be sent to the metadata server. If an I/O request has a length that
1165 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1166 * server. If both file size and I/O size are provided, the client SHOULD
1167 * reach or exceed both thresholds before sending its read or write
1168 * requests to the data server.
1169 */
1170static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1171 struct inode *ino, int iomode)
1172{
1173 struct nfs4_threshold *t = ctx->mdsthreshold;
1174 struct nfs_inode *nfsi = NFS_I(ino);
1175 loff_t fsize = i_size_read(ino);
1176 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1177
1178 if (t == NULL)
1179 return ret;
1180
1181 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1182 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1183
1184 switch (iomode) {
1185 case IOMODE_READ:
1186 if (t->bm & THRESHOLD_RD) {
1187 dprintk("%s fsize %llu\n", __func__, fsize);
1188 size_set = true;
1189 if (fsize < t->rd_sz)
1190 size = true;
1191 }
1192 if (t->bm & THRESHOLD_RD_IO) {
1193 dprintk("%s nfsi->read_io %llu\n", __func__,
1194 nfsi->read_io);
1195 io_set = true;
1196 if (nfsi->read_io < t->rd_io_sz)
1197 io = true;
1198 }
1199 break;
1200 case IOMODE_RW:
1201 if (t->bm & THRESHOLD_WR) {
1202 dprintk("%s fsize %llu\n", __func__, fsize);
1203 size_set = true;
1204 if (fsize < t->wr_sz)
1205 size = true;
1206 }
1207 if (t->bm & THRESHOLD_WR_IO) {
1208 dprintk("%s nfsi->write_io %llu\n", __func__,
1209 nfsi->write_io);
1210 io_set = true;
1211 if (nfsi->write_io < t->wr_io_sz)
1212 io = true;
1213 }
1214 break;
1215 }
1216 if (size_set && io_set) {
1217 if (size && io)
1218 ret = true;
1219 } else if (size || io)
1220 ret = true;
1221
1222 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1223 return ret;
1224}
1225
1226/*
Benny Halevye5e94012010-10-20 00:18:01 -04001227 * Layout segment is retreived from the server if not cached.
1228 * The appropriate layout segment is referenced and returned to the caller.
1229 */
Andy Adamson7c24d942011-06-13 18:22:38 -04001230struct pnfs_layout_segment *
Benny Halevye5e94012010-10-20 00:18:01 -04001231pnfs_update_layout(struct inode *ino,
1232 struct nfs_open_context *ctx,
Benny Halevyfb3296e2011-05-22 19:47:26 +03001233 loff_t pos,
1234 u64 count,
Trond Myklebusta75b9df2011-05-11 18:00:51 -04001235 enum pnfs_iomode iomode,
1236 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001237{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001238 struct pnfs_layout_range arg = {
1239 .iomode = iomode,
1240 .offset = pos,
1241 .length = count,
1242 };
Benny Halevy707ed5f2011-05-22 19:47:46 +03001243 unsigned pg_offset;
Weston Andros Adamson6382a442011-06-01 16:44:44 -04001244 struct nfs_server *server = NFS_SERVER(ino);
1245 struct nfs_client *clp = server->nfs_client;
Benny Halevye5e94012010-10-20 00:18:01 -04001246 struct pnfs_layout_hdr *lo;
1247 struct pnfs_layout_segment *lseg = NULL;
Weston Andros Adamson30005122013-02-28 20:30:10 -05001248 bool first;
Benny Halevye5e94012010-10-20 00:18:01 -04001249
1250 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001251 goto out;
Andy Adamsond23d61c2012-05-23 05:02:37 -04001252
1253 if (pnfs_within_mdsthreshold(ctx, ino, iomode))
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001254 goto out;
Andy Adamsond23d61c2012-05-23 05:02:37 -04001255
Benny Halevye5e94012010-10-20 00:18:01 -04001256 spin_lock(&ino->i_lock);
Peng Tao9fa40752011-07-30 20:52:32 -04001257 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001258 if (lo == NULL) {
1259 spin_unlock(&ino->i_lock);
1260 goto out;
1261 }
Benny Halevye5e94012010-10-20 00:18:01 -04001262
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001263 /* Do we even need to bother with this? */
Trond Myklebusta59c30a2012-03-01 11:17:47 -05001264 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001265 dprintk("%s matches recall, use MDS\n", __func__);
Benny Halevye5e94012010-10-20 00:18:01 -04001266 goto out_unlock;
1267 }
1268
1269 /* if LAYOUTGET already failed once we don't try again */
Trond Myklebustb9e028f2012-09-18 16:41:18 -04001270 if (pnfs_layout_io_test_failed(lo, iomode))
Benny Halevye5e94012010-10-20 00:18:01 -04001271 goto out_unlock;
1272
Andy Adamson568e8c42011-03-01 01:34:22 +00001273 /* Check to see if the layout for the given range already exists */
Benny Halevyfb3296e2011-05-22 19:47:26 +03001274 lseg = pnfs_find_lseg(lo, &arg);
Andy Adamson568e8c42011-03-01 01:34:22 +00001275 if (lseg)
1276 goto out_unlock;
1277
Trond Myklebust19c54ab2012-10-05 16:56:58 -07001278 if (pnfs_layoutgets_blocked(lo, 0))
Fred Isamancf7d63f2011-01-06 11:36:25 +00001279 goto out_unlock;
1280 atomic_inc(&lo->plh_outstanding);
1281
Weston Andros Adamson30005122013-02-28 20:30:10 -05001282 first = list_empty(&lo->plh_layouts) ? true : false;
Fred Isamanf49f9ba2011-02-03 18:28:52 +00001283 spin_unlock(&ino->i_lock);
Weston Andros Adamson30005122013-02-28 20:30:10 -05001284
Fred Isamanf49f9ba2011-02-03 18:28:52 +00001285 if (first) {
Fred Isaman2130ff62011-01-06 11:36:26 +00001286 /* The lo must be on the clp list if there is any
1287 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1288 */
1289 spin_lock(&clp->cl_lock);
Weston Andros Adamson6382a442011-06-01 16:44:44 -04001290 list_add_tail(&lo->plh_layouts, &server->layouts);
Fred Isaman2130ff62011-01-06 11:36:26 +00001291 spin_unlock(&clp->cl_lock);
1292 }
Benny Halevye5e94012010-10-20 00:18:01 -04001293
Benny Halevy707ed5f2011-05-22 19:47:46 +03001294 pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1295 if (pg_offset) {
1296 arg.offset -= pg_offset;
1297 arg.length += pg_offset;
1298 }
Andy Adamson7c24d942011-06-13 18:22:38 -04001299 if (arg.length != NFS4_MAX_UINT64)
1300 arg.length = PAGE_CACHE_ALIGN(arg.length);
Benny Halevy707ed5f2011-05-22 19:47:46 +03001301
Benny Halevyfb3296e2011-05-22 19:47:26 +03001302 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
Fred Isamancf7d63f2011-01-06 11:36:25 +00001303 atomic_dec(&lo->plh_outstanding);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001304out_put_layout_hdr:
Trond Myklebust70c3bd22012-09-18 20:51:13 -04001305 pnfs_put_layout_hdr(lo);
Benny Halevye5e94012010-10-20 00:18:01 -04001306out:
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001307 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1308 "(%s, offset: %llu, length: %llu)\n",
1309 __func__, ino->i_sb->s_id,
1310 (unsigned long long)NFS_FILEID(ino),
1311 lseg == NULL ? "not found" : "found",
1312 iomode==IOMODE_RW ? "read/write" : "read-only",
1313 (unsigned long long)pos,
1314 (unsigned long long)count);
Benny Halevye5e94012010-10-20 00:18:01 -04001315 return lseg;
1316out_unlock:
1317 spin_unlock(&ino->i_lock);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001318 goto out_put_layout_hdr;
Benny Halevye5e94012010-10-20 00:18:01 -04001319}
Andy Adamson7c24d942011-06-13 18:22:38 -04001320EXPORT_SYMBOL_GPL(pnfs_update_layout);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001321
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001322struct pnfs_layout_segment *
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001323pnfs_layout_process(struct nfs4_layoutget *lgp)
1324{
1325 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1326 struct nfs4_layoutget_res *res = &lgp->res;
1327 struct pnfs_layout_segment *lseg;
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001328 struct inode *ino = lo->plh_inode;
Trond Myklebust78096cc2014-02-12 10:02:27 -05001329 LIST_HEAD(free_me);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001330 int status = 0;
1331
1332 /* Inject layout blob into I/O device driver */
Trond Myklebusta75b9df2011-05-11 18:00:51 -04001333 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001334 if (!lseg || IS_ERR(lseg)) {
1335 if (!lseg)
1336 status = -ENOMEM;
1337 else
1338 status = PTR_ERR(lseg);
1339 dprintk("%s: Could not allocate layout: error %d\n",
1340 __func__, status);
1341 goto out;
1342 }
1343
1344 spin_lock(&ino->i_lock);
Trond Myklebusta59c30a2012-03-01 11:17:47 -05001345 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001346 dprintk("%s forget reply due to recall\n", __func__);
1347 goto out_forget_reply;
1348 }
1349
Trond Myklebust19c54ab2012-10-05 16:56:58 -07001350 if (pnfs_layoutgets_blocked(lo, 1) ||
1351 pnfs_layout_stateid_blocked(lo, &res->stateid)) {
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001352 dprintk("%s forget reply due to state\n", __func__);
1353 goto out_forget_reply;
1354 }
Trond Myklebust038d6492012-10-02 16:38:41 -07001355
Trond Myklebust78096cc2014-02-12 10:02:27 -05001356 /* Check that the new stateid matches the old stateid */
1357 pnfs_verify_layout_stateid(lo, &res->stateid, &free_me);
Trond Myklebust038d6492012-10-02 16:38:41 -07001358 /* Done processing layoutget. Set the layout stateid */
1359 pnfs_set_layout_stateid(lo, &res->stateid, false);
1360
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001361 init_lseg(lo, lseg);
Fred Isaman566052c2011-01-06 11:36:20 +00001362 lseg->pls_range = res->range;
Trond Myklebust9369a432012-09-18 20:57:08 -04001363 pnfs_get_lseg(lseg);
Trond Myklebust57036a32012-09-20 16:33:30 -04001364 pnfs_layout_insert_lseg(lo, lseg);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001365
Fred Isamanf7e89172011-01-06 11:36:32 +00001366 if (res->return_on_close) {
1367 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1368 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1369 }
1370
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001371 spin_unlock(&ino->i_lock);
Trond Myklebust78096cc2014-02-12 10:02:27 -05001372 pnfs_free_lseg_list(&free_me);
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001373 return lseg;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001374out:
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001375 return ERR_PTR(status);
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001376
1377out_forget_reply:
1378 spin_unlock(&ino->i_lock);
1379 lseg->pls_layout = lo;
1380 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1381 goto out;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001382}
1383
Trond Myklebustd8007d42011-06-10 13:30:23 -04001384void
1385pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1386{
Peng Tao1fd937b2012-09-25 14:55:57 +08001387 u64 rd_size = req->wb_bytes;
1388
Trond Myklebustbc5a89b2012-10-15 14:58:04 -04001389 WARN_ON_ONCE(pgio->pg_lseg != NULL);
Trond Myklebustd8007d42011-06-10 13:30:23 -04001390
Fred Isaman1825a0d2012-04-20 19:55:31 -04001391 if (req->wb_offset != req->wb_pgbase) {
1392 nfs_pageio_reset_read_mds(pgio);
1393 return;
1394 }
Peng Tao1fd937b2012-09-25 14:55:57 +08001395
1396 if (pgio->pg_dreq == NULL)
1397 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1398 else
1399 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1400
Trond Myklebustd8007d42011-06-10 13:30:23 -04001401 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1402 req->wb_context,
1403 req_offset(req),
Peng Tao1fd937b2012-09-25 14:55:57 +08001404 rd_size,
Trond Myklebustd8007d42011-06-10 13:30:23 -04001405 IOMODE_READ,
1406 GFP_KERNEL);
Trond Myklebuste885de12011-06-10 13:30:23 -04001407 /* If no lseg, fall back to read through mds */
1408 if (pgio->pg_lseg == NULL)
Trond Myklebust1f945352011-07-13 15:59:57 -04001409 nfs_pageio_reset_read_mds(pgio);
Trond Myklebuste885de12011-06-10 13:30:23 -04001410
Trond Myklebustd8007d42011-06-10 13:30:23 -04001411}
1412EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1413
1414void
Peng Tao62965562012-09-25 14:55:57 +08001415pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1416 struct nfs_page *req, u64 wb_size)
Trond Myklebustd8007d42011-06-10 13:30:23 -04001417{
Trond Myklebustbc5a89b2012-10-15 14:58:04 -04001418 WARN_ON_ONCE(pgio->pg_lseg != NULL);
Trond Myklebustd8007d42011-06-10 13:30:23 -04001419
Fred Isaman1825a0d2012-04-20 19:55:31 -04001420 if (req->wb_offset != req->wb_pgbase) {
1421 nfs_pageio_reset_write_mds(pgio);
1422 return;
1423 }
Peng Tao62965562012-09-25 14:55:57 +08001424
Trond Myklebustd8007d42011-06-10 13:30:23 -04001425 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1426 req->wb_context,
1427 req_offset(req),
Peng Tao62965562012-09-25 14:55:57 +08001428 wb_size,
Trond Myklebustd8007d42011-06-10 13:30:23 -04001429 IOMODE_RW,
1430 GFP_NOFS);
Trond Myklebuste885de12011-06-10 13:30:23 -04001431 /* If no lseg, fall back to write through mds */
1432 if (pgio->pg_lseg == NULL)
Trond Myklebust1f945352011-07-13 15:59:57 -04001433 nfs_pageio_reset_write_mds(pgio);
Trond Myklebustd8007d42011-06-10 13:30:23 -04001434}
1435EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1436
Bryan Schumaker1abb50882012-06-20 15:53:47 -04001437void
Fred Isaman061ae2e2012-04-20 14:47:48 -04001438pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1439 const struct nfs_pgio_completion_ops *compl_ops)
Trond Myklebust1751c362011-06-10 13:30:23 -04001440{
1441 struct nfs_server *server = NFS_SERVER(inode);
1442 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1443
1444 if (ld == NULL)
Bryan Schumaker1abb50882012-06-20 15:53:47 -04001445 nfs_pageio_init_read(pgio, inode, compl_ops);
1446 else
1447 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
Trond Myklebust1751c362011-06-10 13:30:23 -04001448}
1449
Bryan Schumaker57208fa2012-06-20 15:53:48 -04001450void
Fred Isaman061ae2e2012-04-20 14:47:48 -04001451pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1452 int ioflags,
1453 const struct nfs_pgio_completion_ops *compl_ops)
Trond Myklebust1751c362011-06-10 13:30:23 -04001454{
1455 struct nfs_server *server = NFS_SERVER(inode);
1456 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1457
1458 if (ld == NULL)
Bryan Schumaker57208fa2012-06-20 15:53:48 -04001459 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1460 else
1461 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
Trond Myklebust1751c362011-06-10 13:30:23 -04001462}
1463
1464bool
Benny Halevydfed2062011-05-25 20:25:22 +03001465pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1466 struct nfs_page *req)
Fred Isamanbae724e2011-03-01 01:34:15 +00001467{
Trond Myklebustd8007d42011-06-10 13:30:23 -04001468 if (pgio->pg_lseg == NULL)
1469 return nfs_generic_pg_test(pgio, prev, req);
Benny Halevy89a58e32011-05-25 20:54:40 +03001470
Trond Myklebust19982ba2011-06-10 13:30:23 -04001471 /*
1472 * Test if a nfs_page is fully contained in the pnfs_layout_range.
1473 * Note that this test makes several assumptions:
1474 * - that the previous nfs_page in the struct nfs_pageio_descriptor
1475 * is known to lie within the range.
1476 * - that the nfs_page being tested is known to be contiguous with the
1477 * previous nfs_page.
1478 * - Layout ranges are page aligned, so we only have to test the
1479 * start offset of the request.
1480 *
1481 * Please also note that 'end_offset' is actually the offset of the
1482 * first byte that lies outside the pnfs_layout_range. FIXME?
1483 *
1484 */
1485 return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1486 pgio->pg_lseg->pls_range.length);
Fred Isamanbae724e2011-03-01 01:34:15 +00001487}
Benny Halevy89a58e32011-05-25 20:54:40 +03001488EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
Fred Isamanbae724e2011-03-01 01:34:15 +00001489
Andy Adamsone7dd79af2012-04-27 17:53:46 -04001490int pnfs_write_done_resend_to_mds(struct inode *inode,
Fred Isaman061ae2e2012-04-20 14:47:48 -04001491 struct list_head *head,
Benny Halevy78f33272013-02-24 09:55:57 -05001492 const struct nfs_pgio_completion_ops *compl_ops,
1493 struct nfs_direct_req *dreq)
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001494{
1495 struct nfs_pageio_descriptor pgio;
1496 LIST_HEAD(failed);
1497
1498 /* Resend all requests through the MDS */
Bryan Schumaker57208fa2012-06-20 15:53:48 -04001499 nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
Benny Halevy78f33272013-02-24 09:55:57 -05001500 pgio.pg_dreq = dreq;
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001501 while (!list_empty(head)) {
1502 struct nfs_page *req = nfs_list_entry(head->next);
1503
1504 nfs_list_remove_request(req);
1505 if (!nfs_pageio_add_request(&pgio, req))
1506 nfs_list_add_request(req, &failed);
1507 }
1508 nfs_pageio_complete(&pgio);
1509
1510 if (!list_empty(&failed)) {
1511 /* For some reason our attempt to resend pages. Mark the
1512 * overall send request as having failed, and let
1513 * nfs_writeback_release_full deal with the error.
1514 */
1515 list_move(&failed, head);
1516 return -EIO;
1517 }
1518 return 0;
1519}
Andy Adamsone7dd79af2012-04-27 17:53:46 -04001520EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001521
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001522static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1523{
Fred Isamancd841602012-04-20 14:47:44 -04001524 struct nfs_pgio_header *hdr = data->header;
1525
1526 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1527 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001528 PNFS_LAYOUTRET_ON_ERROR) {
Fred Isamancd841602012-04-20 14:47:44 -04001529 pnfs_return_layout(hdr->inode);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001530 }
Fred Isaman6c75dc02012-04-20 14:47:47 -04001531 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1532 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
Fred Isaman061ae2e2012-04-20 14:47:48 -04001533 &hdr->pages,
Benny Halevy78f33272013-02-24 09:55:57 -05001534 hdr->completion_ops,
1535 hdr->dreq);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001536}
1537
Benny Halevyd20581a2011-05-22 19:52:03 +03001538/*
1539 * Called by non rpc-based layout drivers
1540 */
Peng Tao8ce160c2011-09-22 21:50:14 -04001541void pnfs_ld_write_done(struct nfs_write_data *data)
Fred Isaman94ad1c82011-03-01 01:34:14 +00001542{
Fred Isamancd841602012-04-20 14:47:44 -04001543 struct nfs_pgio_header *hdr = data->header;
1544
Trond Myklebustcc668ab2013-08-14 15:31:28 -04001545 trace_nfs4_pnfs_write(data, hdr->pnfs_error);
Fred Isamancd841602012-04-20 14:47:44 -04001546 if (!hdr->pnfs_error) {
Benny Halevyd20581a2011-05-22 19:52:03 +03001547 pnfs_set_layoutcommit(data);
Fred Isamancd841602012-04-20 14:47:44 -04001548 hdr->mds_ops->rpc_call_done(&data->task, data);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001549 } else
1550 pnfs_ld_handle_write_error(data);
Fred Isamancd841602012-04-20 14:47:44 -04001551 hdr->mds_ops->rpc_release(data);
Fred Isaman44b83792011-03-03 15:13:44 +00001552}
Benny Halevyd20581a2011-05-22 19:52:03 +03001553EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
Fred Isaman44b83792011-03-03 15:13:44 +00001554
Trond Myklebustdce81292011-07-13 15:59:19 -04001555static void
1556pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1557 struct nfs_write_data *data)
1558{
Fred Isamancd841602012-04-20 14:47:44 -04001559 struct nfs_pgio_header *hdr = data->header;
1560
Fred Isaman6c75dc02012-04-20 14:47:47 -04001561 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1562 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1563 nfs_pageio_reset_write_mds(desc);
1564 desc->pg_recoalesce = 1;
1565 }
Trond Myklebustdce81292011-07-13 15:59:19 -04001566 nfs_writedata_release(data);
1567}
1568
1569static enum pnfs_try_status
Andy Adamson0382b742011-03-03 15:13:45 +00001570pnfs_try_to_write_data(struct nfs_write_data *wdata,
Trond Myklebustdce81292011-07-13 15:59:19 -04001571 const struct rpc_call_ops *call_ops,
1572 struct pnfs_layout_segment *lseg,
1573 int how)
Andy Adamson0382b742011-03-03 15:13:45 +00001574{
Fred Isamancd841602012-04-20 14:47:44 -04001575 struct nfs_pgio_header *hdr = wdata->header;
1576 struct inode *inode = hdr->inode;
Andy Adamson0382b742011-03-03 15:13:45 +00001577 enum pnfs_try_status trypnfs;
1578 struct nfs_server *nfss = NFS_SERVER(inode);
1579
Fred Isamancd841602012-04-20 14:47:44 -04001580 hdr->mds_ops = call_ops;
Andy Adamson0382b742011-03-03 15:13:45 +00001581
1582 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1583 inode->i_ino, wdata->args.count, wdata->args.offset, how);
Andy Adamson0382b742011-03-03 15:13:45 +00001584 trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001585 if (trypnfs != PNFS_NOT_ATTEMPTED)
Andy Adamson0382b742011-03-03 15:13:45 +00001586 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
Andy Adamson0382b742011-03-03 15:13:45 +00001587 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1588 return trypnfs;
1589}
1590
Trond Myklebustdce81292011-07-13 15:59:19 -04001591static void
1592pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1593{
1594 struct nfs_write_data *data;
1595 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1596 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1597
1598 desc->pg_lseg = NULL;
1599 while (!list_empty(head)) {
1600 enum pnfs_try_status trypnfs;
1601
Fred Isaman6c75dc02012-04-20 14:47:47 -04001602 data = list_first_entry(head, struct nfs_write_data, list);
Trond Myklebustdce81292011-07-13 15:59:19 -04001603 list_del_init(&data->list);
1604
1605 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1606 if (trypnfs == PNFS_NOT_ATTEMPTED)
1607 pnfs_write_through_mds(desc, data);
1608 }
Trond Myklebust9369a432012-09-18 20:57:08 -04001609 pnfs_put_lseg(lseg);
Trond Myklebustdce81292011-07-13 15:59:19 -04001610}
1611
Fred Isaman6c75dc02012-04-20 14:47:47 -04001612static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1613{
Trond Myklebust9369a432012-09-18 20:57:08 -04001614 pnfs_put_lseg(hdr->lseg);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001615 nfs_writehdr_free(hdr);
1616}
Bryan Schumaker89d77c82012-07-30 16:05:25 -04001617EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001618
Trond Myklebustdce81292011-07-13 15:59:19 -04001619int
1620pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1621{
Fred Isaman6c75dc02012-04-20 14:47:47 -04001622 struct nfs_write_header *whdr;
1623 struct nfs_pgio_header *hdr;
Trond Myklebustdce81292011-07-13 15:59:19 -04001624 int ret;
1625
Fred Isaman6c75dc02012-04-20 14:47:47 -04001626 whdr = nfs_writehdr_alloc();
1627 if (!whdr) {
Trond Myklebust9b5415b2012-04-27 14:31:47 -04001628 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
Trond Myklebust9369a432012-09-18 20:57:08 -04001629 pnfs_put_lseg(desc->pg_lseg);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001630 desc->pg_lseg = NULL;
1631 return -ENOMEM;
1632 }
1633 hdr = &whdr->header;
1634 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
Trond Myklebust9369a432012-09-18 20:57:08 -04001635 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001636 atomic_inc(&hdr->refcnt);
1637 ret = nfs_generic_flush(desc, hdr);
Trond Myklebustdce81292011-07-13 15:59:19 -04001638 if (ret != 0) {
Trond Myklebust9369a432012-09-18 20:57:08 -04001639 pnfs_put_lseg(desc->pg_lseg);
Trond Myklebustdce81292011-07-13 15:59:19 -04001640 desc->pg_lseg = NULL;
Fred Isaman6c75dc02012-04-20 14:47:47 -04001641 } else
1642 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1643 if (atomic_dec_and_test(&hdr->refcnt))
Fred Isaman061ae2e2012-04-20 14:47:48 -04001644 hdr->completion_ops->completion(hdr);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001645 return ret;
Trond Myklebustdce81292011-07-13 15:59:19 -04001646}
1647EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1648
Andy Adamsone7dd79af2012-04-27 17:53:46 -04001649int pnfs_read_done_resend_to_mds(struct inode *inode,
Fred Isaman061ae2e2012-04-20 14:47:48 -04001650 struct list_head *head,
Benny Halevy78f33272013-02-24 09:55:57 -05001651 const struct nfs_pgio_completion_ops *compl_ops,
1652 struct nfs_direct_req *dreq)
Trond Myklebust62e4a762011-11-10 14:30:37 -05001653{
1654 struct nfs_pageio_descriptor pgio;
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001655 LIST_HEAD(failed);
Trond Myklebust62e4a762011-11-10 14:30:37 -05001656
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001657 /* Resend all requests through the MDS */
Bryan Schumaker1abb50882012-06-20 15:53:47 -04001658 nfs_pageio_init_read(&pgio, inode, compl_ops);
Benny Halevy78f33272013-02-24 09:55:57 -05001659 pgio.pg_dreq = dreq;
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001660 while (!list_empty(head)) {
1661 struct nfs_page *req = nfs_list_entry(head->next);
Trond Myklebust62e4a762011-11-10 14:30:37 -05001662
1663 nfs_list_remove_request(req);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001664 if (!nfs_pageio_add_request(&pgio, req))
1665 nfs_list_add_request(req, &failed);
Trond Myklebust62e4a762011-11-10 14:30:37 -05001666 }
1667 nfs_pageio_complete(&pgio);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001668
1669 if (!list_empty(&failed)) {
1670 list_move(&failed, head);
1671 return -EIO;
1672 }
1673 return 0;
1674}
Andy Adamsone7dd79af2012-04-27 17:53:46 -04001675EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001676
1677static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1678{
Fred Isamancd841602012-04-20 14:47:44 -04001679 struct nfs_pgio_header *hdr = data->header;
1680
1681 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1682 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001683 PNFS_LAYOUTRET_ON_ERROR) {
Fred Isamancd841602012-04-20 14:47:44 -04001684 pnfs_return_layout(hdr->inode);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001685 }
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001686 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1687 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
Fred Isaman061ae2e2012-04-20 14:47:48 -04001688 &hdr->pages,
Benny Halevy78f33272013-02-24 09:55:57 -05001689 hdr->completion_ops,
1690 hdr->dreq);
Trond Myklebust62e4a762011-11-10 14:30:37 -05001691}
1692
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001693/*
Benny Halevyd20581a2011-05-22 19:52:03 +03001694 * Called by non rpc-based layout drivers
1695 */
Peng Tao9b7eecd2011-09-22 21:50:15 -04001696void pnfs_ld_read_done(struct nfs_read_data *data)
Benny Halevyd20581a2011-05-22 19:52:03 +03001697{
Fred Isamancd841602012-04-20 14:47:44 -04001698 struct nfs_pgio_header *hdr = data->header;
1699
Trond Myklebustcc668ab2013-08-14 15:31:28 -04001700 trace_nfs4_pnfs_read(data, hdr->pnfs_error);
Fred Isamancd841602012-04-20 14:47:44 -04001701 if (likely(!hdr->pnfs_error)) {
Benny Halevyd20581a2011-05-22 19:52:03 +03001702 __nfs4_read_done_cb(data);
Fred Isamancd841602012-04-20 14:47:44 -04001703 hdr->mds_ops->rpc_call_done(&data->task, data);
Trond Myklebust62e4a762011-11-10 14:30:37 -05001704 } else
1705 pnfs_ld_handle_read_error(data);
Fred Isamancd841602012-04-20 14:47:44 -04001706 hdr->mds_ops->rpc_release(data);
Benny Halevyd20581a2011-05-22 19:52:03 +03001707}
1708EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1709
Trond Myklebust493292d2011-07-13 15:58:28 -04001710static void
1711pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1712 struct nfs_read_data *data)
1713{
Fred Isamancd841602012-04-20 14:47:44 -04001714 struct nfs_pgio_header *hdr = data->header;
1715
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001716 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1717 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1718 nfs_pageio_reset_read_mds(desc);
1719 desc->pg_recoalesce = 1;
1720 }
Trond Myklebust493292d2011-07-13 15:58:28 -04001721 nfs_readdata_release(data);
1722}
1723
Benny Halevyd20581a2011-05-22 19:52:03 +03001724/*
Andy Adamson64419a92011-03-01 01:34:16 +00001725 * Call the appropriate parallel I/O subsystem read function.
1726 */
Trond Myklebust493292d2011-07-13 15:58:28 -04001727static enum pnfs_try_status
Andy Adamson64419a92011-03-01 01:34:16 +00001728pnfs_try_to_read_data(struct nfs_read_data *rdata,
Trond Myklebust493292d2011-07-13 15:58:28 -04001729 const struct rpc_call_ops *call_ops,
1730 struct pnfs_layout_segment *lseg)
Andy Adamson64419a92011-03-01 01:34:16 +00001731{
Fred Isamancd841602012-04-20 14:47:44 -04001732 struct nfs_pgio_header *hdr = rdata->header;
1733 struct inode *inode = hdr->inode;
Andy Adamson64419a92011-03-01 01:34:16 +00001734 struct nfs_server *nfss = NFS_SERVER(inode);
1735 enum pnfs_try_status trypnfs;
1736
Fred Isamancd841602012-04-20 14:47:44 -04001737 hdr->mds_ops = call_ops;
Andy Adamson64419a92011-03-01 01:34:16 +00001738
1739 dprintk("%s: Reading ino:%lu %u@%llu\n",
1740 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1741
1742 trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001743 if (trypnfs != PNFS_NOT_ATTEMPTED)
Andy Adamson64419a92011-03-01 01:34:16 +00001744 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
Andy Adamson64419a92011-03-01 01:34:16 +00001745 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1746 return trypnfs;
1747}
Andy Adamson863a3c62011-03-23 13:27:54 +00001748
Trond Myklebust493292d2011-07-13 15:58:28 -04001749static void
1750pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1751{
1752 struct nfs_read_data *data;
1753 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1754 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1755
1756 desc->pg_lseg = NULL;
1757 while (!list_empty(head)) {
1758 enum pnfs_try_status trypnfs;
1759
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001760 data = list_first_entry(head, struct nfs_read_data, list);
Trond Myklebust493292d2011-07-13 15:58:28 -04001761 list_del_init(&data->list);
1762
1763 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1764 if (trypnfs == PNFS_NOT_ATTEMPTED)
1765 pnfs_read_through_mds(desc, data);
1766 }
Trond Myklebust9369a432012-09-18 20:57:08 -04001767 pnfs_put_lseg(lseg);
Trond Myklebust493292d2011-07-13 15:58:28 -04001768}
1769
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001770static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1771{
Trond Myklebust9369a432012-09-18 20:57:08 -04001772 pnfs_put_lseg(hdr->lseg);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001773 nfs_readhdr_free(hdr);
1774}
Bryan Schumaker89d77c82012-07-30 16:05:25 -04001775EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001776
Trond Myklebust493292d2011-07-13 15:58:28 -04001777int
1778pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1779{
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001780 struct nfs_read_header *rhdr;
1781 struct nfs_pgio_header *hdr;
Trond Myklebust493292d2011-07-13 15:58:28 -04001782 int ret;
1783
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001784 rhdr = nfs_readhdr_alloc();
1785 if (!rhdr) {
Fred Isaman061ae2e2012-04-20 14:47:48 -04001786 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001787 ret = -ENOMEM;
Trond Myklebust9369a432012-09-18 20:57:08 -04001788 pnfs_put_lseg(desc->pg_lseg);
Trond Myklebust493292d2011-07-13 15:58:28 -04001789 desc->pg_lseg = NULL;
1790 return ret;
1791 }
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001792 hdr = &rhdr->header;
1793 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
Trond Myklebust9369a432012-09-18 20:57:08 -04001794 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001795 atomic_inc(&hdr->refcnt);
1796 ret = nfs_generic_pagein(desc, hdr);
1797 if (ret != 0) {
Trond Myklebust9369a432012-09-18 20:57:08 -04001798 pnfs_put_lseg(desc->pg_lseg);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001799 desc->pg_lseg = NULL;
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001800 } else
1801 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1802 if (atomic_dec_and_test(&hdr->refcnt))
Fred Isaman061ae2e2012-04-20 14:47:48 -04001803 hdr->completion_ops->completion(hdr);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001804 return ret;
Trond Myklebust493292d2011-07-13 15:58:28 -04001805}
1806EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1807
Trond Myklebust71244d92014-01-13 13:34:36 -05001808static void pnfs_clear_layoutcommitting(struct inode *inode)
1809{
1810 unsigned long *bitlock = &NFS_I(inode)->flags;
1811
1812 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
Peter Zijlstra4e857c52014-03-17 18:06:10 +01001813 smp_mb__after_atomic();
Trond Myklebust71244d92014-01-13 13:34:36 -05001814 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1815}
1816
Andy Adamson863a3c62011-03-23 13:27:54 +00001817/*
Peng Taoa9bae562011-07-30 20:52:33 -04001818 * There can be multiple RW segments.
Andy Adamson863a3c62011-03-23 13:27:54 +00001819 */
Peng Taoa9bae562011-07-30 20:52:33 -04001820static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
Andy Adamson863a3c62011-03-23 13:27:54 +00001821{
Peng Taoa9bae562011-07-30 20:52:33 -04001822 struct pnfs_layout_segment *lseg;
Andy Adamson863a3c62011-03-23 13:27:54 +00001823
Peng Taoa9bae562011-07-30 20:52:33 -04001824 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1825 if (lseg->pls_range.iomode == IOMODE_RW &&
Trond Myklebusta073dbf2013-03-20 12:34:32 -04001826 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
Peng Taoa9bae562011-07-30 20:52:33 -04001827 list_add(&lseg->pls_lc_list, listp);
1828 }
Andy Adamson863a3c62011-03-23 13:27:54 +00001829}
1830
Trond Myklebusta073dbf2013-03-20 12:34:32 -04001831static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1832{
1833 struct pnfs_layout_segment *lseg, *tmp;
Trond Myklebusta073dbf2013-03-20 12:34:32 -04001834
1835 /* Matched by references in pnfs_set_layoutcommit */
1836 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1837 list_del_init(&lseg->pls_lc_list);
1838 pnfs_put_lseg(lseg);
1839 }
1840
Trond Myklebust71244d92014-01-13 13:34:36 -05001841 pnfs_clear_layoutcommitting(inode);
Trond Myklebusta073dbf2013-03-20 12:34:32 -04001842}
1843
Peng Tao1b0ae062011-09-22 21:50:12 -04001844void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1845{
Trond Myklebustb9e028f2012-09-18 16:41:18 -04001846 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
Peng Tao1b0ae062011-09-22 21:50:12 -04001847}
1848EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1849
Andy Adamson863a3c62011-03-23 13:27:54 +00001850void
1851pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1852{
Fred Isamancd841602012-04-20 14:47:44 -04001853 struct nfs_pgio_header *hdr = wdata->header;
1854 struct inode *inode = hdr->inode;
1855 struct nfs_inode *nfsi = NFS_I(inode);
Vitaliy Gusev4b8ee2b2011-05-20 01:34:46 +04001856 loff_t end_pos = wdata->mds_offset + wdata->res.count;
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04001857 bool mark_as_dirty = false;
Andy Adamson863a3c62011-03-23 13:27:54 +00001858
Fred Isamancd841602012-04-20 14:47:44 -04001859 spin_lock(&inode->i_lock);
Andy Adamson863a3c62011-03-23 13:27:54 +00001860 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04001861 mark_as_dirty = true;
Andy Adamson863a3c62011-03-23 13:27:54 +00001862 dprintk("%s: Set layoutcommit for inode %lu ",
Fred Isamancd841602012-04-20 14:47:44 -04001863 __func__, inode->i_ino);
Andy Adamson863a3c62011-03-23 13:27:54 +00001864 }
Fred Isamancd841602012-04-20 14:47:44 -04001865 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
Peng Taoa9bae562011-07-30 20:52:33 -04001866 /* references matched in nfs4_layoutcommit_release */
Trond Myklebust9369a432012-09-18 20:57:08 -04001867 pnfs_get_lseg(hdr->lseg);
Peng Taoa9bae562011-07-30 20:52:33 -04001868 }
Peng Taoacff58802011-07-30 20:52:31 -04001869 if (end_pos > nfsi->layout->plh_lwb)
1870 nfsi->layout->plh_lwb = end_pos;
Fred Isamancd841602012-04-20 14:47:44 -04001871 spin_unlock(&inode->i_lock);
Peng Taoacff58802011-07-30 20:52:31 -04001872 dprintk("%s: lseg %p end_pos %llu\n",
Fred Isamancd841602012-04-20 14:47:44 -04001873 __func__, hdr->lseg, nfsi->layout->plh_lwb);
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04001874
1875 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1876 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1877 if (mark_as_dirty)
Fred Isamancd841602012-04-20 14:47:44 -04001878 mark_inode_dirty_sync(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00001879}
1880EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1881
Andy Adamsondb29c082011-07-30 20:52:38 -04001882void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1883{
1884 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1885
1886 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1887 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
Trond Myklebusta073dbf2013-03-20 12:34:32 -04001888 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
Andy Adamsondb29c082011-07-30 20:52:38 -04001889}
1890
Andy Adamsonde4b15c2011-03-12 02:58:09 -05001891/*
1892 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1893 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1894 * data to disk to allow the server to recover the data if it crashes.
1895 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1896 * is off, and a COMMIT is sent to a data server, or
1897 * if WRITEs to a data server return NFS_DATA_SYNC.
1898 */
Andy Adamson863a3c62011-03-23 13:27:54 +00001899int
Andy Adamsonef311532011-03-12 02:58:10 -05001900pnfs_layoutcommit_inode(struct inode *inode, bool sync)
Andy Adamson863a3c62011-03-23 13:27:54 +00001901{
1902 struct nfs4_layoutcommit_data *data;
1903 struct nfs_inode *nfsi = NFS_I(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00001904 loff_t end_pos;
Trond Myklebust71244d92014-01-13 13:34:36 -05001905 int status;
1906
1907 if (!pnfs_layoutcommit_outstanding(inode))
1908 return 0;
Andy Adamson863a3c62011-03-23 13:27:54 +00001909
1910 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1911
Trond Myklebust71244d92014-01-13 13:34:36 -05001912 status = -EAGAIN;
1913 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1914 if (!sync)
1915 goto out;
1916 status = wait_on_bit_lock(&nfsi->flags,
1917 NFS_INO_LAYOUTCOMMITTING,
1918 nfs_wait_bit_killable,
1919 TASK_KILLABLE);
1920 if (status)
1921 goto out;
1922 }
Andy Adamsonde4b15c2011-03-12 02:58:09 -05001923
Trond Myklebust71244d92014-01-13 13:34:36 -05001924 status = -ENOMEM;
Andy Adamson863a3c62011-03-23 13:27:54 +00001925 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1926 data = kzalloc(sizeof(*data), GFP_NOFS);
Trond Myklebust71244d92014-01-13 13:34:36 -05001927 if (!data)
1928 goto clear_layoutcommitting;
Andy Adamson863a3c62011-03-23 13:27:54 +00001929
Trond Myklebust71244d92014-01-13 13:34:36 -05001930 status = 0;
1931 spin_lock(&inode->i_lock);
1932 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1933 goto out_unlock;
Peng Tao92407e72011-10-23 20:21:17 -07001934
Peng Taoa9bae562011-07-30 20:52:33 -04001935 INIT_LIST_HEAD(&data->lseg_list);
Peng Taoa9bae562011-07-30 20:52:33 -04001936 pnfs_list_write_lseg(inode, &data->lseg_list);
Andy Adamson863a3c62011-03-23 13:27:54 +00001937
Peng Taoacff58802011-07-30 20:52:31 -04001938 end_pos = nfsi->layout->plh_lwb;
Peng Taoacff58802011-07-30 20:52:31 -04001939 nfsi->layout->plh_lwb = 0;
Andy Adamson863a3c62011-03-23 13:27:54 +00001940
Trond Myklebustf597c532012-03-04 18:13:56 -05001941 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
Andy Adamson863a3c62011-03-23 13:27:54 +00001942 spin_unlock(&inode->i_lock);
1943
1944 data->args.inode = inode;
Peng Tao9fa40752011-07-30 20:52:32 -04001945 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
Andy Adamson863a3c62011-03-23 13:27:54 +00001946 nfs_fattr_init(&data->fattr);
1947 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1948 data->res.fattr = &data->fattr;
1949 data->args.lastbytewritten = end_pos - 1;
1950 data->res.server = NFS_SERVER(inode);
1951
1952 status = nfs4_proc_layoutcommit(data, sync);
1953out:
Peng Tao92407e72011-10-23 20:21:17 -07001954 if (status)
1955 mark_inode_dirty_sync(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00001956 dprintk("<-- %s status %d\n", __func__, status);
1957 return status;
Trond Myklebust71244d92014-01-13 13:34:36 -05001958out_unlock:
1959 spin_unlock(&inode->i_lock);
Peng Tao92407e72011-10-23 20:21:17 -07001960 kfree(data);
Trond Myklebust71244d92014-01-13 13:34:36 -05001961clear_layoutcommitting:
1962 pnfs_clear_layoutcommitting(inode);
Peng Tao92407e72011-10-23 20:21:17 -07001963 goto out;
Andy Adamson863a3c62011-03-23 13:27:54 +00001964}
Andy Adamson82be4172012-05-23 05:02:35 -04001965
1966struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1967{
1968 struct nfs4_threshold *thp;
1969
1970 thp = kzalloc(sizeof(*thp), GFP_NOFS);
1971 if (!thp) {
1972 dprintk("%s mdsthreshold allocation failed\n", __func__);
1973 return NULL;
1974 }
1975 return thp;
1976}