blob: 0a0e209e8262eab99a5eb9f5e61c253ebac4baf2 [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
Peng Taoaa1e0e3a2014-09-06 00:53:25 +080053static int
54pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
Peng Tao6c166052014-11-17 09:30:40 +080055 enum pnfs_iomode iomode, bool sync);
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
244 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
Peng Tao566f8732014-10-10 23:25:46 +0800245 if (!list_empty(&lo->plh_segs))
246 WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
Trond Myklebust6622c3e2012-09-20 17:23:11 -0400247 pnfs_detach_layout_hdr(lo);
Fred Isamancc6e5342011-01-06 11:36:28 +0000248 spin_unlock(&inode->i_lock);
Trond Myklebust6622c3e2012-09-20 17:23:11 -0400249 pnfs_free_layout_hdr(lo);
Fred Isamancc6e5342011-01-06 11:36:28 +0000250 }
Andy Adamson974cec82010-10-20 00:18:02 -0400251}
252
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400253static int
254pnfs_iomode_to_fail_bit(u32 iomode)
255{
256 return iomode == IOMODE_RW ?
257 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
258}
259
260static void
Trond Myklebust3e621212012-09-24 13:07:16 -0400261pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400262{
Trond Myklebust25c75332012-09-18 17:01:12 -0400263 lo->plh_retry_timestamp = jiffies;
Yanchuan Nian39e88fc2013-01-04 20:19:49 +0800264 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
Trond Myklebust3e621212012-09-24 13:07:16 -0400265 atomic_inc(&lo->plh_refcount);
266}
267
268static void
269pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
270{
271 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
272 atomic_dec(&lo->plh_refcount);
273}
274
275static void
276pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
277{
278 struct inode *inode = lo->plh_inode;
Trond Myklebust115ce572012-09-20 21:19:43 -0400279 struct pnfs_layout_range range = {
280 .iomode = iomode,
281 .offset = 0,
282 .length = NFS4_MAX_UINT64,
283 };
284 LIST_HEAD(head);
Trond Myklebust3e621212012-09-24 13:07:16 -0400285
286 spin_lock(&inode->i_lock);
287 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
Trond Myklebust115ce572012-09-20 21:19:43 -0400288 pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
Trond Myklebust3e621212012-09-24 13:07:16 -0400289 spin_unlock(&inode->i_lock);
Trond Myklebust115ce572012-09-20 21:19:43 -0400290 pnfs_free_lseg_list(&head);
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400291 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
292 iomode == IOMODE_RW ? "RW" : "READ");
293}
294
295static bool
296pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
297{
Trond Myklebust25c75332012-09-18 17:01:12 -0400298 unsigned long start, end;
Trond Myklebust3e621212012-09-24 13:07:16 -0400299 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
300
301 if (test_bit(fail_bit, &lo->plh_flags) == 0)
Trond Myklebust25c75332012-09-18 17:01:12 -0400302 return false;
303 end = jiffies;
304 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
305 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
306 /* It is time to retry the failed layoutgets */
Trond Myklebust3e621212012-09-24 13:07:16 -0400307 pnfs_layout_clear_fail_bit(lo, fail_bit);
Trond Myklebust25c75332012-09-18 17:01:12 -0400308 return false;
309 }
310 return true;
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400311}
312
Andy Adamson974cec82010-10-20 00:18:02 -0400313static void
314init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
315{
Fred Isaman566052c2011-01-06 11:36:20 +0000316 INIT_LIST_HEAD(&lseg->pls_list);
Peng Taoa9bae562011-07-30 20:52:33 -0400317 INIT_LIST_HEAD(&lseg->pls_lc_list);
Fred Isaman4541d162011-01-06 11:36:23 +0000318 atomic_set(&lseg->pls_refcount, 1);
319 smp_mb();
320 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
Fred Isaman566052c2011-01-06 11:36:20 +0000321 lseg->pls_layout = lo;
Andy Adamson974cec82010-10-20 00:18:02 -0400322}
323
Trond Myklebust905ca192012-09-20 20:46:49 -0400324static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
Andy Adamson974cec82010-10-20 00:18:02 -0400325{
Fred Isamanb7edfaa2011-01-06 11:36:21 +0000326 struct inode *ino = lseg->pls_layout->plh_inode;
Andy Adamson974cec82010-10-20 00:18:02 -0400327
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400328 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400329}
330
Fred Isamand684d2a2011-03-01 01:34:13 +0000331static void
Trond Myklebust57036a32012-09-20 16:33:30 -0400332pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
333 struct pnfs_layout_segment *lseg)
Andy Adamson974cec82010-10-20 00:18:02 -0400334{
Trond Myklebust57036a32012-09-20 16:33:30 -0400335 struct inode *inode = lo->plh_inode;
Fred Isamand684d2a2011-03-01 01:34:13 +0000336
Benny Halevyd20581a2011-05-22 19:52:03 +0300337 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
Fred Isamand684d2a2011-03-01 01:34:13 +0000338 list_del_init(&lseg->pls_list);
Trond Myklebust8f0d27d2012-09-20 20:57:11 -0400339 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
340 atomic_dec(&lo->plh_refcount);
Trond Myklebust173f77e2012-09-21 15:49:42 -0400341 if (list_empty(&lo->plh_segs))
342 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
Fred Isamand684d2a2011-03-01 01:34:13 +0000343 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
344}
345
Peng Taoaa1e0e3a2014-09-06 00:53:25 +0800346/* Return true if layoutreturn is needed */
347static bool
348pnfs_layout_need_return(struct pnfs_layout_hdr *lo,
349 struct pnfs_layout_segment *lseg,
350 nfs4_stateid *stateid, enum pnfs_iomode *iomode)
351{
352 struct pnfs_layout_segment *s;
353
354 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
355 return false;
356
357 list_for_each_entry(s, &lo->plh_segs, pls_list)
358 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
359 return false;
360
361 *stateid = lo->plh_stateid;
362 *iomode = lo->plh_return_iomode;
363 /* decreased in pnfs_send_layoutreturn() */
364 lo->plh_block_lgets++;
365 lo->plh_return_iomode = 0;
366 return true;
367}
368
Fred Isamanbae724e2011-03-01 01:34:15 +0000369void
Trond Myklebust9369a432012-09-18 20:57:08 -0400370pnfs_put_lseg(struct pnfs_layout_segment *lseg)
Fred Isamand684d2a2011-03-01 01:34:13 +0000371{
Trond Myklebust57036a32012-09-20 16:33:30 -0400372 struct pnfs_layout_hdr *lo;
Fred Isamand684d2a2011-03-01 01:34:13 +0000373 struct inode *inode;
374
375 if (!lseg)
376 return;
377
Fred Isaman4541d162011-01-06 11:36:23 +0000378 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
379 atomic_read(&lseg->pls_refcount),
380 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
Trond Myklebust57036a32012-09-20 16:33:30 -0400381 lo = lseg->pls_layout;
382 inode = lo->plh_inode;
Fred Isamand684d2a2011-03-01 01:34:13 +0000383 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
Peng Taoaa1e0e3a2014-09-06 00:53:25 +0800384 bool need_return;
385 nfs4_stateid stateid;
386 enum pnfs_iomode iomode;
387
Trond Myklebust8f0d27d2012-09-20 20:57:11 -0400388 pnfs_get_layout_hdr(lo);
Trond Myklebust57036a32012-09-20 16:33:30 -0400389 pnfs_layout_remove_lseg(lo, lseg);
Peng Taoaa1e0e3a2014-09-06 00:53:25 +0800390 need_return = pnfs_layout_need_return(lo, lseg,
391 &stateid, &iomode);
Fred Isamand684d2a2011-03-01 01:34:13 +0000392 spin_unlock(&inode->i_lock);
Trond Myklebust905ca192012-09-20 20:46:49 -0400393 pnfs_free_lseg(lseg);
Peng Taoaa1e0e3a2014-09-06 00:53:25 +0800394 if (need_return)
Peng Tao6c166052014-11-17 09:30:40 +0800395 pnfs_send_layoutreturn(lo, stateid, iomode,
396 true);
Peng Taoaa1e0e3a2014-09-06 00:53:25 +0800397 else
398 pnfs_put_layout_hdr(lo);
Fred Isaman4541d162011-01-06 11:36:23 +0000399 }
Andy Adamson974cec82010-10-20 00:18:02 -0400400}
Trond Myklebust9369a432012-09-18 20:57:08 -0400401EXPORT_SYMBOL_GPL(pnfs_put_lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400402
Trond Myklebust6543f802014-10-08 16:39:12 -0400403static void pnfs_free_lseg_async_work(struct work_struct *work)
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400404{
405 struct pnfs_layout_segment *lseg;
Trond Myklebust6543f802014-10-08 16:39:12 -0400406 struct pnfs_layout_hdr *lo;
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400407
408 lseg = container_of(work, struct pnfs_layout_segment, pls_work);
Trond Myklebust6543f802014-10-08 16:39:12 -0400409 lo = lseg->pls_layout;
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400410
Trond Myklebust6543f802014-10-08 16:39:12 -0400411 pnfs_free_lseg(lseg);
412 pnfs_put_layout_hdr(lo);
413}
414
415static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
416{
417 INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
418 schedule_work(&lseg->pls_work);
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400419}
420
421void
Trond Myklebust6543f802014-10-08 16:39:12 -0400422pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400423{
Trond Myklebust6543f802014-10-08 16:39:12 -0400424 if (!lseg)
425 return;
426
427 assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
428
429 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
430 atomic_read(&lseg->pls_refcount),
431 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
432 if (atomic_dec_and_test(&lseg->pls_refcount)) {
433 struct pnfs_layout_hdr *lo = lseg->pls_layout;
434 pnfs_get_layout_hdr(lo);
435 pnfs_layout_remove_lseg(lo, lseg);
436 pnfs_free_lseg_async(lseg);
437 }
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400438}
Trond Myklebust6543f802014-10-08 16:39:12 -0400439EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
Weston Andros Adamsone6cf82d2014-07-17 20:42:18 -0400440
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400441static u64
Benny Halevyfb3296e2011-05-22 19:47:26 +0300442end_offset(u64 start, u64 len)
Fred Isaman4541d162011-01-06 11:36:23 +0000443{
Benny Halevyfb3296e2011-05-22 19:47:26 +0300444 u64 end;
445
446 end = start + len;
447 return end >= start ? end : NFS4_MAX_UINT64;
448}
449
Benny Halevyfb3296e2011-05-22 19:47:26 +0300450/*
451 * is l2 fully contained in l1?
452 * start1 end1
453 * [----------------------------------)
454 * start2 end2
455 * [----------------)
456 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400457static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400458pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400459 const struct pnfs_layout_range *l2)
Benny Halevyfb3296e2011-05-22 19:47:26 +0300460{
461 u64 start1 = l1->offset;
462 u64 end1 = end_offset(start1, l1->length);
463 u64 start2 = l2->offset;
464 u64 end2 = end_offset(start2, l2->length);
465
466 return (start1 <= start2) && (end1 >= end2);
467}
468
469/*
470 * is l1 and l2 intersecting?
471 * start1 end1
472 * [----------------------------------)
473 * start2 end2
474 * [----------------)
475 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400476static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400477pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400478 const struct pnfs_layout_range *l2)
Benny Halevyfb3296e2011-05-22 19:47:26 +0300479{
480 u64 start1 = l1->offset;
481 u64 end1 = end_offset(start1, l1->length);
482 u64 start2 = l2->offset;
483 u64 end2 = end_offset(start2, l2->length);
484
485 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
486 (end2 == NFS4_MAX_UINT64 || end2 > start1);
487}
488
Fred Isaman4541d162011-01-06 11:36:23 +0000489static bool
Trond Myklebust3cb2df12013-06-03 11:24:36 -0400490should_free_lseg(const struct pnfs_layout_range *lseg_range,
491 const struct pnfs_layout_range *recall_range)
Fred Isaman4541d162011-01-06 11:36:23 +0000492{
Benny Halevy778b5502011-05-22 19:48:02 +0300493 return (recall_range->iomode == IOMODE_ANY ||
494 lseg_range->iomode == recall_range->iomode) &&
Trond Myklebust7dc0ac72013-06-03 11:30:24 -0400495 pnfs_lseg_range_intersecting(lseg_range, recall_range);
Fred Isaman4541d162011-01-06 11:36:23 +0000496}
497
Trond Myklebust24956802013-03-20 13:03:00 -0400498static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
499 struct list_head *tmp_list)
500{
501 if (!atomic_dec_and_test(&lseg->pls_refcount))
502 return false;
503 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
504 list_add(&lseg->pls_list, tmp_list);
505 return true;
506}
507
Fred Isaman4541d162011-01-06 11:36:23 +0000508/* Returns 1 if lseg is removed from list, 0 otherwise */
509static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
510 struct list_head *tmp_list)
511{
512 int rv = 0;
513
514 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
515 /* Remove the reference keeping the lseg in the
516 * list. It will now be removed when all
517 * outstanding io is finished.
518 */
Fred Isamand684d2a2011-03-01 01:34:13 +0000519 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
520 atomic_read(&lseg->pls_refcount));
Trond Myklebust24956802013-03-20 13:03:00 -0400521 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
Fred Isamand684d2a2011-03-01 01:34:13 +0000522 rv = 1;
Fred Isaman4541d162011-01-06 11:36:23 +0000523 }
524 return rv;
525}
526
527/* Returns count of number of matching invalid lsegs remaining in list
528 * after call.
529 */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000530int
Trond Myklebust49a85062012-09-18 20:43:31 -0400531pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
Fred Isaman4541d162011-01-06 11:36:23 +0000532 struct list_head *tmp_list,
Benny Halevy778b5502011-05-22 19:48:02 +0300533 struct pnfs_layout_range *recall_range)
Andy Adamson974cec82010-10-20 00:18:02 -0400534{
535 struct pnfs_layout_segment *lseg, *next;
Fred Isaman4541d162011-01-06 11:36:23 +0000536 int invalid = 0, removed = 0;
Andy Adamson974cec82010-10-20 00:18:02 -0400537
538 dprintk("%s:Begin lo %p\n", __func__, lo);
539
Trond Myklebust8006bfb2012-09-21 14:48:04 -0400540 if (list_empty(&lo->plh_segs))
Fred Isaman38511722011-02-03 18:28:50 +0000541 return 0;
Fred Isaman4541d162011-01-06 11:36:23 +0000542 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
Benny Halevy778b5502011-05-22 19:48:02 +0300543 if (!recall_range ||
544 should_free_lseg(&lseg->pls_range, recall_range)) {
Fred Isaman4541d162011-01-06 11:36:23 +0000545 dprintk("%s: freeing lseg %p iomode %d "
546 "offset %llu length %llu\n", __func__,
547 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
548 lseg->pls_range.length);
549 invalid++;
550 removed += mark_lseg_invalid(lseg, tmp_list);
551 }
552 dprintk("%s:Return %i\n", __func__, invalid - removed);
553 return invalid - removed;
Andy Adamson974cec82010-10-20 00:18:02 -0400554}
555
Fred Isamanf49f9ba2011-02-03 18:28:52 +0000556/* note free_me must contain lsegs from a single layout_hdr */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000557void
Fred Isaman4541d162011-01-06 11:36:23 +0000558pnfs_free_lseg_list(struct list_head *free_me)
Andy Adamson974cec82010-10-20 00:18:02 -0400559{
Fred Isaman4541d162011-01-06 11:36:23 +0000560 struct pnfs_layout_segment *lseg, *tmp;
Andy Adamson974cec82010-10-20 00:18:02 -0400561
Fred Isamanf49f9ba2011-02-03 18:28:52 +0000562 if (list_empty(free_me))
563 return;
564
Fred Isaman4541d162011-01-06 11:36:23 +0000565 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
Fred Isaman566052c2011-01-06 11:36:20 +0000566 list_del(&lseg->pls_list);
Trond Myklebust905ca192012-09-20 20:46:49 -0400567 pnfs_free_lseg(lseg);
Andy Adamson974cec82010-10-20 00:18:02 -0400568 }
569}
570
Benny Halevye5e94012010-10-20 00:18:01 -0400571void
572pnfs_destroy_layout(struct nfs_inode *nfsi)
573{
574 struct pnfs_layout_hdr *lo;
Andy Adamson974cec82010-10-20 00:18:02 -0400575 LIST_HEAD(tmp_list);
Benny Halevye5e94012010-10-20 00:18:01 -0400576
577 spin_lock(&nfsi->vfs_inode.i_lock);
578 lo = nfsi->layout;
579 if (lo) {
Fred Isaman38511722011-02-03 18:28:50 +0000580 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
Trond Myklebust49a85062012-09-18 20:43:31 -0400581 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
Trond Myklebust3e621212012-09-24 13:07:16 -0400582 pnfs_get_layout_hdr(lo);
583 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
584 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
585 spin_unlock(&nfsi->vfs_inode.i_lock);
586 pnfs_free_lseg_list(&tmp_list);
587 pnfs_put_layout_hdr(lo);
588 } else
589 spin_unlock(&nfsi->vfs_inode.i_lock);
Benny Halevye5e94012010-10-20 00:18:01 -0400590}
Andy Adamson041245c2012-04-27 17:53:53 -0400591EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
Benny Halevye5e94012010-10-20 00:18:01 -0400592
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500593static bool
594pnfs_layout_add_bulk_destroy_list(struct inode *inode,
595 struct list_head *layout_list)
596{
597 struct pnfs_layout_hdr *lo;
598 bool ret = false;
599
600 spin_lock(&inode->i_lock);
601 lo = NFS_I(inode)->layout;
602 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
603 pnfs_get_layout_hdr(lo);
604 list_add(&lo->plh_bulk_destroy, layout_list);
605 ret = true;
606 }
607 spin_unlock(&inode->i_lock);
608 return ret;
609}
610
611/* Caller must hold rcu_read_lock and clp->cl_lock */
612static int
613pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
614 struct nfs_server *server,
615 struct list_head *layout_list)
616{
617 struct pnfs_layout_hdr *lo, *next;
618 struct inode *inode;
619
620 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
621 inode = igrab(lo->plh_inode);
622 if (inode == NULL)
623 continue;
624 list_del_init(&lo->plh_layouts);
625 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
626 continue;
627 rcu_read_unlock();
628 spin_unlock(&clp->cl_lock);
629 iput(inode);
630 spin_lock(&clp->cl_lock);
631 rcu_read_lock();
632 return -EAGAIN;
633 }
634 return 0;
635}
636
637static int
638pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
639 bool is_bulk_recall)
640{
641 struct pnfs_layout_hdr *lo;
642 struct inode *inode;
643 struct pnfs_layout_range range = {
644 .iomode = IOMODE_ANY,
645 .offset = 0,
646 .length = NFS4_MAX_UINT64,
647 };
648 LIST_HEAD(lseg_list);
649 int ret = 0;
650
651 while (!list_empty(layout_list)) {
652 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
653 plh_bulk_destroy);
654 dprintk("%s freeing layout for inode %lu\n", __func__,
655 lo->plh_inode->i_ino);
656 inode = lo->plh_inode;
Christoph Hellwig7c5d1872014-09-10 08:23:29 -0700657
658 pnfs_layoutcommit_inode(inode, false);
659
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500660 spin_lock(&inode->i_lock);
661 list_del_init(&lo->plh_bulk_destroy);
662 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
663 if (is_bulk_recall)
664 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
665 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
666 ret = -EAGAIN;
667 spin_unlock(&inode->i_lock);
668 pnfs_free_lseg_list(&lseg_list);
669 pnfs_put_layout_hdr(lo);
670 iput(inode);
671 }
672 return ret;
673}
674
675int
676pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
677 struct nfs_fsid *fsid,
678 bool is_recall)
679{
680 struct nfs_server *server;
681 LIST_HEAD(layout_list);
682
683 spin_lock(&clp->cl_lock);
684 rcu_read_lock();
685restart:
686 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
687 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
688 continue;
689 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
690 server,
691 &layout_list) != 0)
692 goto restart;
693 }
694 rcu_read_unlock();
695 spin_unlock(&clp->cl_lock);
696
697 if (list_empty(&layout_list))
698 return 0;
699 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
700}
701
702int
703pnfs_destroy_layouts_byclid(struct nfs_client *clp,
704 bool is_recall)
705{
706 struct nfs_server *server;
707 LIST_HEAD(layout_list);
708
709 spin_lock(&clp->cl_lock);
710 rcu_read_lock();
711restart:
712 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
713 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
714 server,
715 &layout_list) != 0)
716 goto restart;
717 }
718 rcu_read_unlock();
719 spin_unlock(&clp->cl_lock);
720
721 if (list_empty(&layout_list))
722 return 0;
723 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
724}
725
Andy Adamson974cec82010-10-20 00:18:02 -0400726/*
727 * Called by the state manger to remove all layouts established under an
728 * expired lease.
729 */
730void
731pnfs_destroy_all_layouts(struct nfs_client *clp)
732{
Andy Adamsonc47abcf2011-06-15 17:52:40 -0400733 nfs4_deviceid_mark_client_invalid(clp);
734 nfs4_deviceid_purge_client(clp);
735
Trond Myklebustfd9a8d72013-02-12 09:48:42 -0500736 pnfs_destroy_layouts_byclid(clp, false);
Andy Adamson974cec82010-10-20 00:18:02 -0400737}
738
Trond Myklebust5a655032012-10-02 16:47:14 -0700739/*
740 * Compare 2 layout stateid sequence ids, to see which is newer,
741 * taking into account wraparound issues.
742 */
743static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
744{
Trond Myklebust2c64c572014-02-12 09:36:59 -0500745 return (s32)(s1 - s2) > 0;
Trond Myklebust5a655032012-10-02 16:47:14 -0700746}
747
Fred Isamanfd6002e2011-01-06 11:36:22 +0000748/* update lo->plh_stateid with new if is more recent */
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000749void
750pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
751 bool update_barrier)
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400752{
Trond Myklebust22aaf712012-10-04 16:32:22 -0700753 u32 oldseq, newseq, new_barrier;
754 int empty = list_empty(&lo->plh_segs);
Andy Adamson974cec82010-10-20 00:18:02 -0400755
Trond Myklebust2d2f24a2012-03-04 18:13:57 -0500756 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
757 newseq = be32_to_cpu(new->seqid);
Trond Myklebust22aaf712012-10-04 16:32:22 -0700758 if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
Trond Myklebustf597c532012-03-04 18:13:56 -0500759 nfs4_stateid_copy(&lo->plh_stateid, new);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000760 if (update_barrier) {
Trond Myklebust22aaf712012-10-04 16:32:22 -0700761 new_barrier = be32_to_cpu(new->seqid);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000762 } else {
763 /* Because of wraparound, we want to keep the barrier
Trond Myklebust22aaf712012-10-04 16:32:22 -0700764 * "close" to the current seqids.
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000765 */
Trond Myklebust22aaf712012-10-04 16:32:22 -0700766 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000767 }
Trond Myklebust22aaf712012-10-04 16:32:22 -0700768 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
769 lo->plh_barrier = new_barrier;
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000770 }
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400771}
772
Trond Myklebust19c54ab2012-10-05 16:56:58 -0700773static bool
774pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
775 const nfs4_stateid *stateid)
776{
777 u32 seqid = be32_to_cpu(stateid->seqid);
778
779 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
780}
781
Peng Taoce6ab4f2014-09-06 00:53:24 +0800782static bool
783pnfs_layout_returning(const struct pnfs_layout_hdr *lo,
784 struct pnfs_layout_range *range)
785{
786 return test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
787 (lo->plh_return_iomode == IOMODE_ANY ||
788 lo->plh_return_iomode == range->iomode);
789}
790
Fred Isamancf7d63f2011-01-06 11:36:25 +0000791/* lget is set to 1 if called from inside send_layoutget call chain */
792static bool
Peng Taoce6ab4f2014-09-06 00:53:24 +0800793pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo,
794 struct pnfs_layout_range *range, int lget)
Fred Isamancf7d63f2011-01-06 11:36:25 +0000795{
Fred Isamanf7e89172011-01-06 11:36:32 +0000796 return lo->plh_block_lgets ||
797 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
Fred Isaman43f1b3d2011-01-06 11:36:30 +0000798 (list_empty(&lo->plh_segs) &&
Peng Taoce6ab4f2014-09-06 00:53:24 +0800799 (atomic_read(&lo->plh_outstanding) > lget)) ||
800 pnfs_layout_returning(lo, range);
Fred Isamancf7d63f2011-01-06 11:36:25 +0000801}
802
Fred Isamanfd6002e2011-01-06 11:36:22 +0000803int
804pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
Peng Taoce6ab4f2014-09-06 00:53:24 +0800805 struct pnfs_layout_range *range,
Fred Isamanfd6002e2011-01-06 11:36:22 +0000806 struct nfs4_state *open_state)
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400807{
Fred Isamanfd6002e2011-01-06 11:36:22 +0000808 int status = 0;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400809
810 dprintk("--> %s\n", __func__);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000811 spin_lock(&lo->plh_inode->i_lock);
Peng Taoce6ab4f2014-09-06 00:53:24 +0800812 if (pnfs_layoutgets_blocked(lo, range, 1)) {
Fred Isamancf7d63f2011-01-06 11:36:25 +0000813 status = -EAGAIN;
Trond Myklebust5d422302013-03-14 16:57:48 -0400814 } else if (!nfs4_valid_open_stateid(open_state)) {
815 status = -EBADF;
Christoph Hellwig47abade2014-08-21 11:09:22 -0500816 } else if (list_empty(&lo->plh_segs) ||
817 test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
Fred Isamanfd6002e2011-01-06 11:36:22 +0000818 int seq;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400819
Fred Isamanfd6002e2011-01-06 11:36:22 +0000820 do {
821 seq = read_seqbegin(&open_state->seqlock);
Trond Myklebustf597c532012-03-04 18:13:56 -0500822 nfs4_stateid_copy(dst, &open_state->stateid);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000823 } while (read_seqretry(&open_state->seqlock, seq));
824 } else
Trond Myklebustf597c532012-03-04 18:13:56 -0500825 nfs4_stateid_copy(dst, &lo->plh_stateid);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000826 spin_unlock(&lo->plh_inode->i_lock);
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400827 dprintk("<-- %s\n", __func__);
Fred Isamanfd6002e2011-01-06 11:36:22 +0000828 return status;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400829}
830
831/*
832* Get layout from server.
833* for now, assume that whole file layouts are requested.
834* arg->offset: 0
835* arg->length: all ones
836*/
Benny Halevye5e94012010-10-20 00:18:01 -0400837static struct pnfs_layout_segment *
838send_layoutget(struct pnfs_layout_hdr *lo,
839 struct nfs_open_context *ctx,
Benny Halevyfb3296e2011-05-22 19:47:26 +0300840 struct pnfs_layout_range *range,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400841 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -0400842{
Fred Isamanb7edfaa2011-01-06 11:36:21 +0000843 struct inode *ino = lo->plh_inode;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400844 struct nfs_server *server = NFS_SERVER(ino);
845 struct nfs4_layoutget *lgp;
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -0400846 struct pnfs_layout_segment *lseg;
Benny Halevye5e94012010-10-20 00:18:01 -0400847
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400848 dprintk("--> %s\n", __func__);
849
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400850 lgp = kzalloc(sizeof(*lgp), gfp_flags);
Fred Isamancf7d63f2011-01-06 11:36:25 +0000851 if (lgp == NULL)
Andy Adamson974cec82010-10-20 00:18:02 -0400852 return NULL;
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400853
Benny Halevyfb3296e2011-05-22 19:47:26 +0300854 lgp->args.minlength = PAGE_CACHE_SIZE;
855 if (lgp->args.minlength > range->length)
856 lgp->args.minlength = range->length;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400857 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
Benny Halevyfb3296e2011-05-22 19:47:26 +0300858 lgp->args.range = *range;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400859 lgp->args.type = server->pnfs_curr_ld->id;
860 lgp->args.inode = ino;
861 lgp->args.ctx = get_nfs_open_context(ctx);
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400862 lgp->gfp_flags = gfp_flags;
Trond Myklebust6ab59342013-05-20 10:49:34 -0400863 lgp->cred = lo->plh_lc_cred;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400864
865 /* Synchronously retrieve layout information from server and
866 * store in lseg.
867 */
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -0400868 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
869 if (IS_ERR(lseg)) {
870 switch (PTR_ERR(lseg)) {
871 case -ENOMEM:
872 case -ERESTARTSYS:
873 break;
874 default:
875 /* remember that LAYOUTGET failed and suspend trying */
Trond Myklebustb9e028f2012-09-18 16:41:18 -0400876 pnfs_layout_io_set_failed(lo, range->iomode);
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -0400877 }
878 return NULL;
Andy Adamsonb1f69b72010-10-20 00:18:03 -0400879 }
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400880
Andy Adamson974cec82010-10-20 00:18:02 -0400881 return lseg;
882}
883
Trond Myklebust24956802013-03-20 13:03:00 -0400884static void pnfs_clear_layoutcommit(struct inode *inode,
885 struct list_head *head)
886{
887 struct nfs_inode *nfsi = NFS_I(inode);
888 struct pnfs_layout_segment *lseg, *tmp;
889
890 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
891 return;
892 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
893 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
894 continue;
895 pnfs_lseg_dec_and_remove_zero(lseg, head);
896 }
897}
898
Peng Taof40eb5d2014-09-06 00:53:22 +0800899static int
900pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
Peng Tao6c166052014-11-17 09:30:40 +0800901 enum pnfs_iomode iomode, bool sync)
Peng Taof40eb5d2014-09-06 00:53:22 +0800902{
903 struct inode *ino = lo->plh_inode;
904 struct nfs4_layoutreturn *lrp;
905 int status = 0;
906
907 lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
908 if (unlikely(lrp == NULL)) {
909 status = -ENOMEM;
910 spin_lock(&ino->i_lock);
911 lo->plh_block_lgets--;
912 spin_unlock(&ino->i_lock);
913 pnfs_put_layout_hdr(lo);
914 goto out;
915 }
916
917 lrp->args.stateid = stateid;
918 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
919 lrp->args.inode = ino;
Peng Tao15eb67c2014-11-17 09:30:36 +0800920 lrp->args.range.iomode = iomode;
921 lrp->args.range.offset = 0;
922 lrp->args.range.length = NFS4_MAX_UINT64;
Peng Taof40eb5d2014-09-06 00:53:22 +0800923 lrp->args.layout = lo;
924 lrp->clp = NFS_SERVER(ino)->nfs_client;
925 lrp->cred = lo->plh_lc_cred;
926
Peng Tao6c166052014-11-17 09:30:40 +0800927 status = nfs4_proc_layoutreturn(lrp, sync);
Peng Taof40eb5d2014-09-06 00:53:22 +0800928out:
Peng Taoe736a5b2014-09-06 00:53:26 +0800929 if (status) {
930 spin_lock(&ino->i_lock);
931 clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
932 spin_unlock(&ino->i_lock);
933 }
Peng Taof40eb5d2014-09-06 00:53:22 +0800934 dprintk("<-- %s status: %d\n", __func__, status);
935 return status;
936}
937
Andy Adamson293b3b02012-06-20 15:03:34 -0400938/*
939 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
940 * when the layout segment list is empty.
941 *
942 * Note that a pnfs_layout_hdr can exist with an empty layout segment
943 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
944 * deviceid is marked invalid.
945 */
Benny Halevycbe82602011-05-22 19:52:37 +0300946int
947_pnfs_return_layout(struct inode *ino)
948{
949 struct pnfs_layout_hdr *lo = NULL;
950 struct nfs_inode *nfsi = NFS_I(ino);
951 LIST_HEAD(tmp_list);
Benny Halevycbe82602011-05-22 19:52:37 +0300952 nfs4_stateid stateid;
Andy Adamson293b3b02012-06-20 15:03:34 -0400953 int status = 0, empty;
Benny Halevycbe82602011-05-22 19:52:37 +0300954
Andy Adamson366d5052012-06-20 15:03:33 -0400955 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
Benny Halevycbe82602011-05-22 19:52:37 +0300956
957 spin_lock(&ino->i_lock);
958 lo = nfsi->layout;
Trond Myklebuste5929f32012-09-21 16:37:02 -0400959 if (!lo) {
Benny Halevycbe82602011-05-22 19:52:37 +0300960 spin_unlock(&ino->i_lock);
Andy Adamson293b3b02012-06-20 15:03:34 -0400961 dprintk("NFS: %s no layout to return\n", __func__);
962 goto out;
Benny Halevycbe82602011-05-22 19:52:37 +0300963 }
964 stateid = nfsi->layout->plh_stateid;
965 /* Reference matched in nfs4_layoutreturn_release */
Trond Myklebust70c3bd22012-09-18 20:51:13 -0400966 pnfs_get_layout_hdr(lo);
Andy Adamson293b3b02012-06-20 15:03:34 -0400967 empty = list_empty(&lo->plh_segs);
Trond Myklebust24956802013-03-20 13:03:00 -0400968 pnfs_clear_layoutcommit(ino, &tmp_list);
Trond Myklebust49a85062012-09-18 20:43:31 -0400969 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
Christoph Hellwigc88953d2014-09-10 08:23:31 -0700970
971 if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
972 struct pnfs_layout_range range = {
973 .iomode = IOMODE_ANY,
974 .offset = 0,
975 .length = NFS4_MAX_UINT64,
976 };
977 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
978 }
979
Andy Adamson293b3b02012-06-20 15:03:34 -0400980 /* Don't send a LAYOUTRETURN if list was initially empty */
981 if (empty) {
982 spin_unlock(&ino->i_lock);
Trond Myklebust70c3bd22012-09-18 20:51:13 -0400983 pnfs_put_layout_hdr(lo);
Andy Adamson293b3b02012-06-20 15:03:34 -0400984 dprintk("NFS: %s no layout segments to return\n", __func__);
985 goto out;
986 }
Christoph Hellwig47abade2014-08-21 11:09:22 -0500987
988 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
Fred Isamanea0ded72011-06-15 12:31:02 -0400989 lo->plh_block_lgets++;
Benny Halevycbe82602011-05-22 19:52:37 +0300990 spin_unlock(&ino->i_lock);
991 pnfs_free_lseg_list(&tmp_list);
992
Peng Tao6c166052014-11-17 09:30:40 +0800993 status = pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
Benny Halevycbe82602011-05-22 19:52:37 +0300994out:
995 dprintk("<-- %s status: %d\n", __func__, status);
996 return status;
997}
Andy Adamson0a57cda2012-04-27 17:53:50 -0400998EXPORT_SYMBOL_GPL(_pnfs_return_layout);
Benny Halevycbe82602011-05-22 19:52:37 +0300999
Trond Myklebust24028672013-03-20 13:23:33 -04001000int
1001pnfs_commit_and_return_layout(struct inode *inode)
1002{
1003 struct pnfs_layout_hdr *lo;
1004 int ret;
1005
1006 spin_lock(&inode->i_lock);
1007 lo = NFS_I(inode)->layout;
1008 if (lo == NULL) {
1009 spin_unlock(&inode->i_lock);
1010 return 0;
1011 }
1012 pnfs_get_layout_hdr(lo);
1013 /* Block new layoutgets and read/write to ds */
1014 lo->plh_block_lgets++;
1015 spin_unlock(&inode->i_lock);
1016 filemap_fdatawait(inode->i_mapping);
1017 ret = pnfs_layoutcommit_inode(inode, true);
1018 if (ret == 0)
1019 ret = _pnfs_return_layout(inode);
1020 spin_lock(&inode->i_lock);
1021 lo->plh_block_lgets--;
1022 spin_unlock(&inode->i_lock);
1023 pnfs_put_layout_hdr(lo);
1024 return ret;
1025}
1026
Fred Isamanf7e89172011-01-06 11:36:32 +00001027bool pnfs_roc(struct inode *ino)
1028{
1029 struct pnfs_layout_hdr *lo;
1030 struct pnfs_layout_segment *lseg, *tmp;
1031 LIST_HEAD(tmp_list);
1032 bool found = false;
1033
1034 spin_lock(&ino->i_lock);
1035 lo = NFS_I(ino)->layout;
1036 if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
1037 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1038 goto out_nolayout;
1039 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1040 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1041 mark_lseg_invalid(lseg, &tmp_list);
1042 found = true;
1043 }
1044 if (!found)
1045 goto out_nolayout;
1046 lo->plh_block_lgets++;
Trond Myklebust70c3bd22012-09-18 20:51:13 -04001047 pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
Fred Isamanf7e89172011-01-06 11:36:32 +00001048 spin_unlock(&ino->i_lock);
1049 pnfs_free_lseg_list(&tmp_list);
1050 return true;
1051
1052out_nolayout:
1053 spin_unlock(&ino->i_lock);
1054 return false;
1055}
1056
1057void pnfs_roc_release(struct inode *ino)
1058{
1059 struct pnfs_layout_hdr *lo;
1060
1061 spin_lock(&ino->i_lock);
1062 lo = NFS_I(ino)->layout;
1063 lo->plh_block_lgets--;
Trond Myklebust6622c3e2012-09-20 17:23:11 -04001064 if (atomic_dec_and_test(&lo->plh_refcount)) {
1065 pnfs_detach_layout_hdr(lo);
1066 spin_unlock(&ino->i_lock);
1067 pnfs_free_layout_hdr(lo);
1068 } else
1069 spin_unlock(&ino->i_lock);
Fred Isamanf7e89172011-01-06 11:36:32 +00001070}
1071
1072void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1073{
1074 struct pnfs_layout_hdr *lo;
1075
1076 spin_lock(&ino->i_lock);
1077 lo = NFS_I(ino)->layout;
Trond Myklebust0f35ad62012-10-04 16:28:17 -07001078 if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
Fred Isamanf7e89172011-01-06 11:36:32 +00001079 lo->plh_barrier = barrier;
1080 spin_unlock(&ino->i_lock);
1081}
1082
Trond Myklebust7fdab062012-09-20 20:15:57 -04001083bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
Fred Isamanf7e89172011-01-06 11:36:32 +00001084{
1085 struct nfs_inode *nfsi = NFS_I(ino);
Trond Myklebust7fdab062012-09-20 20:15:57 -04001086 struct pnfs_layout_hdr *lo;
Fred Isamanf7e89172011-01-06 11:36:32 +00001087 struct pnfs_layout_segment *lseg;
Trond Myklebust7fdab062012-09-20 20:15:57 -04001088 u32 current_seqid;
Fred Isamanf7e89172011-01-06 11:36:32 +00001089 bool found = false;
1090
1091 spin_lock(&ino->i_lock);
1092 list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
1093 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
Trond Myklebust7fdab062012-09-20 20:15:57 -04001094 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
Fred Isamanf7e89172011-01-06 11:36:32 +00001095 found = true;
Trond Myklebust7fdab062012-09-20 20:15:57 -04001096 goto out;
Fred Isamanf7e89172011-01-06 11:36:32 +00001097 }
Trond Myklebust7fdab062012-09-20 20:15:57 -04001098 lo = nfsi->layout;
1099 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
Fred Isamanf7e89172011-01-06 11:36:32 +00001100
Trond Myklebust7fdab062012-09-20 20:15:57 -04001101 /* Since close does not return a layout stateid for use as
1102 * a barrier, we choose the worst-case barrier.
1103 */
1104 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1105out:
Fred Isamanf7e89172011-01-06 11:36:32 +00001106 spin_unlock(&ino->i_lock);
1107 return found;
1108}
1109
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001110/*
1111 * Compare two layout segments for sorting into layout cache.
1112 * We want to preferentially return RW over RO layouts, so ensure those
1113 * are seen first.
1114 */
1115static s64
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001116pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001117 const struct pnfs_layout_range *l2)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001118{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001119 s64 d;
1120
1121 /* high offset > low offset */
1122 d = l1->offset - l2->offset;
1123 if (d)
1124 return d;
1125
1126 /* short length > long length */
1127 d = l2->length - l1->length;
1128 if (d)
1129 return d;
1130
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001131 /* read > read/write */
Benny Halevyfb3296e2011-05-22 19:47:26 +03001132 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001133}
1134
Andy Adamson974cec82010-10-20 00:18:02 -04001135static void
Trond Myklebust57036a32012-09-20 16:33:30 -04001136pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
Andy Adamson974cec82010-10-20 00:18:02 -04001137 struct pnfs_layout_segment *lseg)
1138{
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001139 struct pnfs_layout_segment *lp;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001140
Andy Adamson974cec82010-10-20 00:18:02 -04001141 dprintk("%s:Begin\n", __func__);
1142
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001143 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001144 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001145 continue;
Fred Isaman566052c2011-01-06 11:36:20 +00001146 list_add_tail(&lseg->pls_list, &lp->pls_list);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001147 dprintk("%s: inserted lseg %p "
1148 "iomode %d offset %llu length %llu before "
1149 "lp %p iomode %d offset %llu length %llu\n",
Fred Isaman566052c2011-01-06 11:36:20 +00001150 __func__, lseg, lseg->pls_range.iomode,
1151 lseg->pls_range.offset, lseg->pls_range.length,
1152 lp, lp->pls_range.iomode, lp->pls_range.offset,
1153 lp->pls_range.length);
Benny Halevyfb3296e2011-05-22 19:47:26 +03001154 goto out;
Andy Adamson974cec82010-10-20 00:18:02 -04001155 }
Benny Halevyfb3296e2011-05-22 19:47:26 +03001156 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1157 dprintk("%s: inserted lseg %p "
1158 "iomode %d offset %llu length %llu at tail\n",
1159 __func__, lseg, lseg->pls_range.iomode,
1160 lseg->pls_range.offset, lseg->pls_range.length);
1161out:
Trond Myklebust70c3bd22012-09-18 20:51:13 -04001162 pnfs_get_layout_hdr(lo);
Andy Adamson974cec82010-10-20 00:18:02 -04001163
1164 dprintk("%s:Return\n", __func__);
Benny Halevye5e94012010-10-20 00:18:01 -04001165}
1166
1167static struct pnfs_layout_hdr *
Peng Tao9fa40752011-07-30 20:52:32 -04001168alloc_init_layout_hdr(struct inode *ino,
1169 struct nfs_open_context *ctx,
1170 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001171{
1172 struct pnfs_layout_hdr *lo;
1173
Benny Halevy636fb9c2011-05-22 19:51:33 +03001174 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
Benny Halevye5e94012010-10-20 00:18:01 -04001175 if (!lo)
1176 return NULL;
Fred Isamancc6e5342011-01-06 11:36:28 +00001177 atomic_set(&lo->plh_refcount, 1);
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001178 INIT_LIST_HEAD(&lo->plh_layouts);
1179 INIT_LIST_HEAD(&lo->plh_segs);
Trond Myklebustfd9a8d72013-02-12 09:48:42 -05001180 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001181 lo->plh_inode = ino;
Trond Myklebust5cc22162013-05-21 09:26:49 -04001182 lo->plh_lc_cred = get_rpccred(ctx->cred);
Benny Halevye5e94012010-10-20 00:18:01 -04001183 return lo;
1184}
1185
1186static struct pnfs_layout_hdr *
Peng Tao9fa40752011-07-30 20:52:32 -04001187pnfs_find_alloc_layout(struct inode *ino,
1188 struct nfs_open_context *ctx,
1189 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001190{
1191 struct nfs_inode *nfsi = NFS_I(ino);
1192 struct pnfs_layout_hdr *new = NULL;
1193
1194 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1195
Trond Myklebust251ec412012-10-02 15:41:05 -07001196 if (nfsi->layout != NULL)
1197 goto out_existing;
Benny Halevye5e94012010-10-20 00:18:01 -04001198 spin_unlock(&ino->i_lock);
Peng Tao9fa40752011-07-30 20:52:32 -04001199 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
Benny Halevye5e94012010-10-20 00:18:01 -04001200 spin_lock(&ino->i_lock);
1201
Trond Myklebust251ec412012-10-02 15:41:05 -07001202 if (likely(nfsi->layout == NULL)) { /* Won the race? */
Benny Halevye5e94012010-10-20 00:18:01 -04001203 nfsi->layout = new;
Trond Myklebust251ec412012-10-02 15:41:05 -07001204 return new;
Yanchuan Nian7175fe92012-10-31 16:05:48 +08001205 } else if (new != NULL)
1206 pnfs_free_layout_hdr(new);
Trond Myklebust251ec412012-10-02 15:41:05 -07001207out_existing:
1208 pnfs_get_layout_hdr(nfsi->layout);
Benny Halevye5e94012010-10-20 00:18:01 -04001209 return nfsi->layout;
1210}
1211
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001212/*
1213 * iomode matching rules:
1214 * iomode lseg match
1215 * ----- ----- -----
1216 * ANY READ true
1217 * ANY RW true
1218 * RW READ false
1219 * RW RW true
1220 * READ READ true
1221 * READ RW true
1222 */
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001223static bool
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001224pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
Trond Myklebust3cb2df12013-06-03 11:24:36 -04001225 const struct pnfs_layout_range *range)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001226{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001227 struct pnfs_layout_range range1;
1228
1229 if ((range->iomode == IOMODE_RW &&
1230 ls_range->iomode != IOMODE_RW) ||
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001231 !pnfs_lseg_range_intersecting(ls_range, range))
Benny Halevyfb3296e2011-05-22 19:47:26 +03001232 return 0;
1233
1234 /* range1 covers only the first byte in the range */
1235 range1 = *range;
1236 range1.length = 1;
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001237 return pnfs_lseg_range_contained(ls_range, &range1);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001238}
1239
1240/*
1241 * lookup range in layout
1242 */
Benny Halevye5e94012010-10-20 00:18:01 -04001243static struct pnfs_layout_segment *
Benny Halevyfb3296e2011-05-22 19:47:26 +03001244pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1245 struct pnfs_layout_range *range)
Benny Halevye5e94012010-10-20 00:18:01 -04001246{
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001247 struct pnfs_layout_segment *lseg, *ret = NULL;
1248
1249 dprintk("%s:Begin\n", __func__);
1250
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001251 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
Fred Isaman4541d162011-01-06 11:36:23 +00001252 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
Peng Taoce6ab4f2014-09-06 00:53:24 +08001253 !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
Trond Myklebust7dc0ac72013-06-03 11:30:24 -04001254 pnfs_lseg_range_match(&lseg->pls_range, range)) {
Trond Myklebust9369a432012-09-18 20:57:08 -04001255 ret = pnfs_get_lseg(lseg);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001256 break;
1257 }
Benny Halevyd771e3a2011-06-14 16:30:16 -04001258 if (lseg->pls_range.offset > range->offset)
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001259 break;
1260 }
1261
1262 dprintk("%s:Return lseg %p ref %d\n",
Fred Isaman4541d162011-01-06 11:36:23 +00001263 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001264 return ret;
Benny Halevye5e94012010-10-20 00:18:01 -04001265}
1266
1267/*
Andy Adamsond23d61c2012-05-23 05:02:37 -04001268 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1269 * to the MDS or over pNFS
1270 *
1271 * The nfs_inode read_io and write_io fields are cumulative counters reset
1272 * when there are no layout segments. Note that in pnfs_update_layout iomode
1273 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1274 * WRITE request.
1275 *
1276 * A return of true means use MDS I/O.
1277 *
1278 * From rfc 5661:
1279 * If a file's size is smaller than the file size threshold, data accesses
1280 * SHOULD be sent to the metadata server. If an I/O request has a length that
1281 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1282 * server. If both file size and I/O size are provided, the client SHOULD
1283 * reach or exceed both thresholds before sending its read or write
1284 * requests to the data server.
1285 */
1286static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1287 struct inode *ino, int iomode)
1288{
1289 struct nfs4_threshold *t = ctx->mdsthreshold;
1290 struct nfs_inode *nfsi = NFS_I(ino);
1291 loff_t fsize = i_size_read(ino);
1292 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1293
1294 if (t == NULL)
1295 return ret;
1296
1297 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1298 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1299
1300 switch (iomode) {
1301 case IOMODE_READ:
1302 if (t->bm & THRESHOLD_RD) {
1303 dprintk("%s fsize %llu\n", __func__, fsize);
1304 size_set = true;
1305 if (fsize < t->rd_sz)
1306 size = true;
1307 }
1308 if (t->bm & THRESHOLD_RD_IO) {
1309 dprintk("%s nfsi->read_io %llu\n", __func__,
1310 nfsi->read_io);
1311 io_set = true;
1312 if (nfsi->read_io < t->rd_io_sz)
1313 io = true;
1314 }
1315 break;
1316 case IOMODE_RW:
1317 if (t->bm & THRESHOLD_WR) {
1318 dprintk("%s fsize %llu\n", __func__, fsize);
1319 size_set = true;
1320 if (fsize < t->wr_sz)
1321 size = true;
1322 }
1323 if (t->bm & THRESHOLD_WR_IO) {
1324 dprintk("%s nfsi->write_io %llu\n", __func__,
1325 nfsi->write_io);
1326 io_set = true;
1327 if (nfsi->write_io < t->wr_io_sz)
1328 io = true;
1329 }
1330 break;
1331 }
1332 if (size_set && io_set) {
1333 if (size && io)
1334 ret = true;
1335 } else if (size || io)
1336 ret = true;
1337
1338 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1339 return ret;
1340}
1341
1342/*
Benny Halevye5e94012010-10-20 00:18:01 -04001343 * Layout segment is retreived from the server if not cached.
1344 * The appropriate layout segment is referenced and returned to the caller.
1345 */
Andy Adamson7c24d942011-06-13 18:22:38 -04001346struct pnfs_layout_segment *
Benny Halevye5e94012010-10-20 00:18:01 -04001347pnfs_update_layout(struct inode *ino,
1348 struct nfs_open_context *ctx,
Benny Halevyfb3296e2011-05-22 19:47:26 +03001349 loff_t pos,
1350 u64 count,
Trond Myklebusta75b9df2011-05-11 18:00:51 -04001351 enum pnfs_iomode iomode,
1352 gfp_t gfp_flags)
Benny Halevye5e94012010-10-20 00:18:01 -04001353{
Benny Halevyfb3296e2011-05-22 19:47:26 +03001354 struct pnfs_layout_range arg = {
1355 .iomode = iomode,
1356 .offset = pos,
1357 .length = count,
1358 };
Benny Halevy707ed5f2011-05-22 19:47:46 +03001359 unsigned pg_offset;
Weston Andros Adamson6382a442011-06-01 16:44:44 -04001360 struct nfs_server *server = NFS_SERVER(ino);
1361 struct nfs_client *clp = server->nfs_client;
Benny Halevye5e94012010-10-20 00:18:01 -04001362 struct pnfs_layout_hdr *lo;
1363 struct pnfs_layout_segment *lseg = NULL;
Peng Tao9bf87482014-08-22 17:37:41 +08001364 bool first;
Benny Halevye5e94012010-10-20 00:18:01 -04001365
1366 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001367 goto out;
Andy Adamsond23d61c2012-05-23 05:02:37 -04001368
1369 if (pnfs_within_mdsthreshold(ctx, ino, iomode))
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001370 goto out;
Andy Adamsond23d61c2012-05-23 05:02:37 -04001371
Peng Tao9bf87482014-08-22 17:37:41 +08001372lookup_again:
1373 first = false;
Benny Halevye5e94012010-10-20 00:18:01 -04001374 spin_lock(&ino->i_lock);
Peng Tao9fa40752011-07-30 20:52:32 -04001375 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001376 if (lo == NULL) {
1377 spin_unlock(&ino->i_lock);
1378 goto out;
1379 }
Benny Halevye5e94012010-10-20 00:18:01 -04001380
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001381 /* Do we even need to bother with this? */
Trond Myklebusta59c30a2012-03-01 11:17:47 -05001382 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001383 dprintk("%s matches recall, use MDS\n", __func__);
Benny Halevye5e94012010-10-20 00:18:01 -04001384 goto out_unlock;
1385 }
1386
1387 /* if LAYOUTGET already failed once we don't try again */
Trond Myklebustb9e028f2012-09-18 16:41:18 -04001388 if (pnfs_layout_io_test_failed(lo, iomode))
Benny Halevye5e94012010-10-20 00:18:01 -04001389 goto out_unlock;
1390
Peng Tao9bf87482014-08-22 17:37:41 +08001391 first = list_empty(&lo->plh_segs);
1392 if (first) {
1393 /* The first layoutget for the file. Need to serialize per
1394 * RFC 5661 Errata 3208.
1395 */
1396 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1397 &lo->plh_flags)) {
1398 spin_unlock(&ino->i_lock);
1399 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1400 TASK_UNINTERRUPTIBLE);
1401 pnfs_put_layout_hdr(lo);
1402 goto lookup_again;
1403 }
1404 } else {
1405 /* Check to see if the layout for the given range
1406 * already exists
1407 */
1408 lseg = pnfs_find_lseg(lo, &arg);
1409 if (lseg)
1410 goto out_unlock;
1411 }
Andy Adamson568e8c42011-03-01 01:34:22 +00001412
Peng Taoce6ab4f2014-09-06 00:53:24 +08001413 if (pnfs_layoutgets_blocked(lo, &arg, 0))
Fred Isamancf7d63f2011-01-06 11:36:25 +00001414 goto out_unlock;
1415 atomic_inc(&lo->plh_outstanding);
Fred Isamanf49f9ba2011-02-03 18:28:52 +00001416 spin_unlock(&ino->i_lock);
Weston Andros Adamson30005122013-02-28 20:30:10 -05001417
Peng Taoabb9a002014-08-22 17:37:40 +08001418 if (list_empty(&lo->plh_layouts)) {
Fred Isaman2130ff62011-01-06 11:36:26 +00001419 /* The lo must be on the clp list if there is any
1420 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1421 */
1422 spin_lock(&clp->cl_lock);
Peng Taoabb9a002014-08-22 17:37:40 +08001423 if (list_empty(&lo->plh_layouts))
1424 list_add_tail(&lo->plh_layouts, &server->layouts);
Fred Isaman2130ff62011-01-06 11:36:26 +00001425 spin_unlock(&clp->cl_lock);
1426 }
Benny Halevye5e94012010-10-20 00:18:01 -04001427
Benny Halevy707ed5f2011-05-22 19:47:46 +03001428 pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1429 if (pg_offset) {
1430 arg.offset -= pg_offset;
1431 arg.length += pg_offset;
1432 }
Andy Adamson7c24d942011-06-13 18:22:38 -04001433 if (arg.length != NFS4_MAX_UINT64)
1434 arg.length = PAGE_CACHE_ALIGN(arg.length);
Benny Halevy707ed5f2011-05-22 19:47:46 +03001435
Benny Halevyfb3296e2011-05-22 19:47:26 +03001436 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
Fred Isamancf7d63f2011-01-06 11:36:25 +00001437 atomic_dec(&lo->plh_outstanding);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001438out_put_layout_hdr:
Peng Tao9bf87482014-08-22 17:37:41 +08001439 if (first) {
1440 unsigned long *bitlock = &lo->plh_flags;
1441
1442 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1443 smp_mb__after_atomic();
1444 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1445 }
Trond Myklebust70c3bd22012-09-18 20:51:13 -04001446 pnfs_put_layout_hdr(lo);
Benny Halevye5e94012010-10-20 00:18:01 -04001447out:
Trond Myklebustf86bbcf2012-09-26 11:21:40 -04001448 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1449 "(%s, offset: %llu, length: %llu)\n",
1450 __func__, ino->i_sb->s_id,
1451 (unsigned long long)NFS_FILEID(ino),
1452 lseg == NULL ? "not found" : "found",
1453 iomode==IOMODE_RW ? "read/write" : "read-only",
1454 (unsigned long long)pos,
1455 (unsigned long long)count);
Benny Halevye5e94012010-10-20 00:18:01 -04001456 return lseg;
1457out_unlock:
1458 spin_unlock(&ino->i_lock);
Trond Myklebust830ffb52012-09-20 21:25:19 -04001459 goto out_put_layout_hdr;
Benny Halevye5e94012010-10-20 00:18:01 -04001460}
Andy Adamson7c24d942011-06-13 18:22:38 -04001461EXPORT_SYMBOL_GPL(pnfs_update_layout);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001462
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001463struct pnfs_layout_segment *
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001464pnfs_layout_process(struct nfs4_layoutget *lgp)
1465{
1466 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1467 struct nfs4_layoutget_res *res = &lgp->res;
1468 struct pnfs_layout_segment *lseg;
Fred Isamanb7edfaa2011-01-06 11:36:21 +00001469 struct inode *ino = lo->plh_inode;
Trond Myklebust78096cc2014-02-12 10:02:27 -05001470 LIST_HEAD(free_me);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001471 int status = 0;
1472
1473 /* Inject layout blob into I/O device driver */
Trond Myklebusta75b9df2011-05-11 18:00:51 -04001474 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001475 if (!lseg || IS_ERR(lseg)) {
1476 if (!lseg)
1477 status = -ENOMEM;
1478 else
1479 status = PTR_ERR(lseg);
1480 dprintk("%s: Could not allocate layout: error %d\n",
1481 __func__, status);
1482 goto out;
1483 }
1484
Christoph Hellwig1013df612014-08-21 11:09:18 -05001485 init_lseg(lo, lseg);
1486 lseg->pls_range = res->range;
1487
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001488 spin_lock(&ino->i_lock);
Trond Myklebusta59c30a2012-03-01 11:17:47 -05001489 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001490 dprintk("%s forget reply due to recall\n", __func__);
1491 goto out_forget_reply;
1492 }
1493
Peng Taoce6ab4f2014-09-06 00:53:24 +08001494 if (pnfs_layoutgets_blocked(lo, &lgp->args.range, 1)) {
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001495 dprintk("%s forget reply due to state\n", __func__);
1496 goto out_forget_reply;
1497 }
Trond Myklebust038d6492012-10-02 16:38:41 -07001498
Christoph Hellwig362f7472014-08-21 11:09:20 -05001499 if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1500 /* existing state ID, make sure the sequence number matches. */
1501 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1502 dprintk("%s forget reply due to sequence\n", __func__);
1503 goto out_forget_reply;
1504 }
1505 pnfs_set_layout_stateid(lo, &res->stateid, false);
1506 } else {
1507 /*
1508 * We got an entirely new state ID. Mark all segments for the
1509 * inode invalid, and don't bother validating the stateid
1510 * sequence number.
1511 */
1512 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1513
1514 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1515 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1516 }
Trond Myklebust038d6492012-10-02 16:38:41 -07001517
Christoph Hellwig47abade2014-08-21 11:09:22 -05001518 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1519
Trond Myklebust9369a432012-09-18 20:57:08 -04001520 pnfs_get_lseg(lseg);
Trond Myklebust57036a32012-09-20 16:33:30 -04001521 pnfs_layout_insert_lseg(lo, lseg);
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001522
Fred Isamanf7e89172011-01-06 11:36:32 +00001523 if (res->return_on_close) {
1524 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1525 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1526 }
1527
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001528 spin_unlock(&ino->i_lock);
Trond Myklebust78096cc2014-02-12 10:02:27 -05001529 pnfs_free_lseg_list(&free_me);
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001530 return lseg;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001531out:
Trond Myklebusta0b0a6e2012-09-17 17:12:15 -04001532 return ERR_PTR(status);
Fred Isaman43f1b3d2011-01-06 11:36:30 +00001533
1534out_forget_reply:
1535 spin_unlock(&ino->i_lock);
1536 lseg->pls_layout = lo;
1537 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1538 goto out;
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001539}
1540
Peng Tao016256d2014-09-06 00:53:23 +08001541static void
1542pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1543 struct list_head *tmp_list,
1544 struct pnfs_layout_range *return_range)
1545{
1546 struct pnfs_layout_segment *lseg, *next;
1547
1548 dprintk("%s:Begin lo %p\n", __func__, lo);
1549
1550 if (list_empty(&lo->plh_segs))
1551 return;
1552
1553 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1554 if (should_free_lseg(&lseg->pls_range, return_range)) {
1555 dprintk("%s: marking lseg %p iomode %d "
1556 "offset %llu length %llu\n", __func__,
1557 lseg, lseg->pls_range.iomode,
1558 lseg->pls_range.offset,
1559 lseg->pls_range.length);
1560 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1561 mark_lseg_invalid(lseg, tmp_list);
1562 }
1563}
1564
1565void pnfs_error_mark_layout_for_return(struct inode *inode,
1566 struct pnfs_layout_segment *lseg)
1567{
1568 struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1569 int iomode = pnfs_iomode_to_fail_bit(lseg->pls_range.iomode);
1570 struct pnfs_layout_range range = {
1571 .iomode = lseg->pls_range.iomode,
1572 .offset = 0,
1573 .length = NFS4_MAX_UINT64,
1574 };
1575 LIST_HEAD(free_me);
1576
1577 spin_lock(&inode->i_lock);
1578 /* set failure bit so that pnfs path will be retried later */
1579 pnfs_layout_set_fail_bit(lo, iomode);
1580 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1581 if (lo->plh_return_iomode == 0)
1582 lo->plh_return_iomode = range.iomode;
1583 else if (lo->plh_return_iomode != range.iomode)
1584 lo->plh_return_iomode = IOMODE_ANY;
1585 /*
1586 * mark all matching lsegs so that we are sure to have no live
1587 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1588 * for how it works.
1589 */
1590 pnfs_mark_matching_lsegs_return(lo, &free_me, &range);
1591 spin_unlock(&inode->i_lock);
1592 pnfs_free_lseg_list(&free_me);
1593}
1594EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1595
Trond Myklebustd8007d42011-06-10 13:30:23 -04001596void
1597pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1598{
Peng Tao1fd937b2012-09-25 14:55:57 +08001599 u64 rd_size = req->wb_bytes;
1600
Trond Myklebustbc5a89b2012-10-15 14:58:04 -04001601 WARN_ON_ONCE(pgio->pg_lseg != NULL);
Trond Myklebustd8007d42011-06-10 13:30:23 -04001602
Peng Tao1fd937b2012-09-25 14:55:57 +08001603 if (pgio->pg_dreq == NULL)
1604 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1605 else
1606 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1607
Trond Myklebustd8007d42011-06-10 13:30:23 -04001608 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1609 req->wb_context,
1610 req_offset(req),
Peng Tao1fd937b2012-09-25 14:55:57 +08001611 rd_size,
Trond Myklebustd8007d42011-06-10 13:30:23 -04001612 IOMODE_READ,
1613 GFP_KERNEL);
Trond Myklebuste885de12011-06-10 13:30:23 -04001614 /* If no lseg, fall back to read through mds */
1615 if (pgio->pg_lseg == NULL)
Trond Myklebust1f945352011-07-13 15:59:57 -04001616 nfs_pageio_reset_read_mds(pgio);
Trond Myklebuste885de12011-06-10 13:30:23 -04001617
Trond Myklebustd8007d42011-06-10 13:30:23 -04001618}
1619EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1620
1621void
Peng Tao62965562012-09-25 14:55:57 +08001622pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1623 struct nfs_page *req, u64 wb_size)
Trond Myklebustd8007d42011-06-10 13:30:23 -04001624{
Trond Myklebustbc5a89b2012-10-15 14:58:04 -04001625 WARN_ON_ONCE(pgio->pg_lseg != NULL);
Trond Myklebustd8007d42011-06-10 13:30:23 -04001626
1627 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1628 req->wb_context,
1629 req_offset(req),
Peng Tao62965562012-09-25 14:55:57 +08001630 wb_size,
Trond Myklebustd8007d42011-06-10 13:30:23 -04001631 IOMODE_RW,
1632 GFP_NOFS);
Trond Myklebuste885de12011-06-10 13:30:23 -04001633 /* If no lseg, fall back to write through mds */
1634 if (pgio->pg_lseg == NULL)
Trond Myklebust1f945352011-07-13 15:59:57 -04001635 nfs_pageio_reset_write_mds(pgio);
Trond Myklebustd8007d42011-06-10 13:30:23 -04001636}
1637EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1638
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04001639void
1640pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1641{
1642 if (desc->pg_lseg) {
1643 pnfs_put_lseg(desc->pg_lseg);
1644 desc->pg_lseg = NULL;
1645 }
1646}
1647EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1648
Weston Andros Adamsonb4fdac12014-05-15 11:56:43 -04001649/*
1650 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1651 * of bytes (maximum @req->wb_bytes) that can be coalesced.
1652 */
1653size_t
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001654pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1655 struct nfs_page *prev, struct nfs_page *req)
Fred Isamanbae724e2011-03-01 01:34:15 +00001656{
Weston Andros Adamson0f9c4292014-05-15 11:56:51 -04001657 unsigned int size;
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001658 u64 seg_end, req_start, seg_left;
Weston Andros Adamson0f9c4292014-05-15 11:56:51 -04001659
1660 size = nfs_generic_pg_test(pgio, prev, req);
Weston Andros Adamson0f9c4292014-05-15 11:56:51 -04001661 if (!size)
1662 return 0;
Benny Halevy89a58e32011-05-25 20:54:40 +03001663
Trond Myklebust19982ba2011-06-10 13:30:23 -04001664 /*
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001665 * 'size' contains the number of bytes left in the current page (up
1666 * to the original size asked for in @req->wb_bytes).
1667 *
1668 * Calculate how many bytes are left in the layout segment
1669 * and if there are less bytes than 'size', return that instead.
Trond Myklebust19982ba2011-06-10 13:30:23 -04001670 *
1671 * Please also note that 'end_offset' is actually the offset of the
1672 * first byte that lies outside the pnfs_layout_range. FIXME?
1673 *
1674 */
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001675 if (pgio->pg_lseg) {
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001676 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1677 pgio->pg_lseg->pls_range.length);
1678 req_start = req_offset(req);
1679 WARN_ON_ONCE(req_start > seg_end);
1680 /* start of request is past the last byte of this segment */
1681 if (req_start >= seg_end)
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001682 return 0;
Weston Andros Adamsonc5e20cb2014-06-09 17:47:26 -04001683
1684 /* adjust 'size' iff there are fewer bytes left in the
1685 * segment than what nfs_generic_pg_test returned */
1686 seg_left = seg_end - req_start;
1687 if (seg_left < size)
1688 size = (unsigned int)seg_left;
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001689 }
Weston Andros Adamson0f9c4292014-05-15 11:56:51 -04001690
Weston Andros Adamson19b54842014-05-15 11:56:55 -04001691 return size;
Fred Isamanbae724e2011-03-01 01:34:15 +00001692}
Benny Halevy89a58e32011-05-25 20:54:40 +03001693EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
Fred Isamanbae724e2011-03-01 01:34:15 +00001694
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001695int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001696{
1697 struct nfs_pageio_descriptor pgio;
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001698
1699 /* Resend all requests through the MDS */
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001700 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1701 hdr->completion_ops);
1702 return nfs_pageio_resend(&pgio, hdr);
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001703}
Andy Adamsone7dd79af2012-04-27 17:53:46 -04001704EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
Trond Myklebuste2fecb22012-01-06 08:57:46 -05001705
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001706static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001707{
Fred Isamancd841602012-04-20 14:47:44 -04001708
1709 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1710 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001711 PNFS_LAYOUTRET_ON_ERROR) {
Fred Isamancd841602012-04-20 14:47:44 -04001712 pnfs_return_layout(hdr->inode);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001713 }
Fred Isaman6c75dc02012-04-20 14:47:47 -04001714 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001715 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001716}
1717
Benny Halevyd20581a2011-05-22 19:52:03 +03001718/*
1719 * Called by non rpc-based layout drivers
1720 */
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001721void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
Fred Isaman94ad1c82011-03-01 01:34:14 +00001722{
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001723 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
Fred Isamancd841602012-04-20 14:47:44 -04001724 if (!hdr->pnfs_error) {
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001725 pnfs_set_layoutcommit(hdr);
1726 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001727 } else
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001728 pnfs_ld_handle_write_error(hdr);
1729 hdr->mds_ops->rpc_release(hdr);
Fred Isaman44b83792011-03-03 15:13:44 +00001730}
Benny Halevyd20581a2011-05-22 19:52:03 +03001731EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
Fred Isaman44b83792011-03-03 15:13:44 +00001732
Trond Myklebustdce81292011-07-13 15:59:19 -04001733static void
1734pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001735 struct nfs_pgio_header *hdr)
Trond Myklebustdce81292011-07-13 15:59:19 -04001736{
Peng Tao48d635f2014-11-10 08:35:35 +08001737 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001738
Fred Isaman6c75dc02012-04-20 14:47:47 -04001739 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001740 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001741 nfs_pageio_reset_write_mds(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001742 mirror->pg_recoalesce = 1;
Fred Isaman6c75dc02012-04-20 14:47:47 -04001743 }
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001744 nfs_pgio_data_destroy(hdr);
Trond Myklebustdce81292011-07-13 15:59:19 -04001745}
1746
1747static enum pnfs_try_status
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001748pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
Trond Myklebustdce81292011-07-13 15:59:19 -04001749 const struct rpc_call_ops *call_ops,
1750 struct pnfs_layout_segment *lseg,
1751 int how)
Andy Adamson0382b742011-03-03 15:13:45 +00001752{
Fred Isamancd841602012-04-20 14:47:44 -04001753 struct inode *inode = hdr->inode;
Andy Adamson0382b742011-03-03 15:13:45 +00001754 enum pnfs_try_status trypnfs;
1755 struct nfs_server *nfss = NFS_SERVER(inode);
1756
Fred Isamancd841602012-04-20 14:47:44 -04001757 hdr->mds_ops = call_ops;
Andy Adamson0382b742011-03-03 15:13:45 +00001758
1759 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001760 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1761 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001762 if (trypnfs != PNFS_NOT_ATTEMPTED)
Andy Adamson0382b742011-03-03 15:13:45 +00001763 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
Andy Adamson0382b742011-03-03 15:13:45 +00001764 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1765 return trypnfs;
1766}
1767
Trond Myklebustdce81292011-07-13 15:59:19 -04001768static void
Weston Andros Adamson7f714722014-05-15 11:56:53 -04001769pnfs_do_write(struct nfs_pageio_descriptor *desc,
1770 struct nfs_pgio_header *hdr, int how)
Trond Myklebustdce81292011-07-13 15:59:19 -04001771{
Trond Myklebustdce81292011-07-13 15:59:19 -04001772 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1773 struct pnfs_layout_segment *lseg = desc->pg_lseg;
Weston Andros Adamson7f714722014-05-15 11:56:53 -04001774 enum pnfs_try_status trypnfs;
Trond Myklebustdce81292011-07-13 15:59:19 -04001775
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001776 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
Weston Andros Adamson7f714722014-05-15 11:56:53 -04001777 if (trypnfs == PNFS_NOT_ATTEMPTED)
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001778 pnfs_write_through_mds(desc, hdr);
Trond Myklebustdce81292011-07-13 15:59:19 -04001779}
1780
Fred Isaman6c75dc02012-04-20 14:47:47 -04001781static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1782{
Trond Myklebust9369a432012-09-18 20:57:08 -04001783 pnfs_put_lseg(hdr->lseg);
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04001784 nfs_pgio_header_free(hdr);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001785}
Bryan Schumaker89d77c82012-07-30 16:05:25 -04001786EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001787
Trond Myklebustdce81292011-07-13 15:59:19 -04001788int
1789pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1790{
Peng Tao48d635f2014-11-10 08:35:35 +08001791 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001792
Fred Isaman6c75dc02012-04-20 14:47:47 -04001793 struct nfs_pgio_header *hdr;
Trond Myklebustdce81292011-07-13 15:59:19 -04001794 int ret;
1795
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04001796 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1797 if (!hdr) {
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001798 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
Fred Isaman6c75dc02012-04-20 14:47:47 -04001799 return -ENOMEM;
1800 }
Fred Isaman6c75dc02012-04-20 14:47:47 -04001801 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04001802
Trond Myklebust9369a432012-09-18 20:57:08 -04001803 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
Anna Schumakeref2c4882014-05-06 09:12:36 -04001804 ret = nfs_generic_pgio(desc, hdr);
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04001805 if (!ret)
Weston Andros Adamson7f714722014-05-15 11:56:53 -04001806 pnfs_do_write(desc, hdr, desc->pg_ioflags);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001807
Fred Isaman6c75dc02012-04-20 14:47:47 -04001808 return ret;
Trond Myklebustdce81292011-07-13 15:59:19 -04001809}
1810EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1811
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001812int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
Trond Myklebust62e4a762011-11-10 14:30:37 -05001813{
1814 struct nfs_pageio_descriptor pgio;
1815
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001816 /* Resend all requests through the MDS */
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001817 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1818 return nfs_pageio_resend(&pgio, hdr);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001819}
Andy Adamsone7dd79af2012-04-27 17:53:46 -04001820EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001821
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001822static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001823{
Fred Isamancd841602012-04-20 14:47:44 -04001824 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1825 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001826 PNFS_LAYOUTRET_ON_ERROR) {
Fred Isamancd841602012-04-20 14:47:44 -04001827 pnfs_return_layout(hdr->inode);
Fred Isaman1acbbb4e12012-04-20 14:47:37 -04001828 }
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001829 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
Weston Andros Adamson53113ad2014-06-09 11:48:38 -04001830 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
Trond Myklebust62e4a762011-11-10 14:30:37 -05001831}
1832
Andy Adamsonb1f69b72010-10-20 00:18:03 -04001833/*
Benny Halevyd20581a2011-05-22 19:52:03 +03001834 * Called by non rpc-based layout drivers
1835 */
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001836void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
Benny Halevyd20581a2011-05-22 19:52:03 +03001837{
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001838 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
Fred Isamancd841602012-04-20 14:47:44 -04001839 if (likely(!hdr->pnfs_error)) {
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001840 __nfs4_read_done_cb(hdr);
1841 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
Trond Myklebust62e4a762011-11-10 14:30:37 -05001842 } else
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001843 pnfs_ld_handle_read_error(hdr);
1844 hdr->mds_ops->rpc_release(hdr);
Benny Halevyd20581a2011-05-22 19:52:03 +03001845}
1846EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1847
Trond Myklebust493292d2011-07-13 15:58:28 -04001848static void
1849pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001850 struct nfs_pgio_header *hdr)
Trond Myklebust493292d2011-07-13 15:58:28 -04001851{
Peng Tao48d635f2014-11-10 08:35:35 +08001852 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001853
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001854 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001855 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001856 nfs_pageio_reset_read_mds(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001857 mirror->pg_recoalesce = 1;
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001858 }
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001859 nfs_pgio_data_destroy(hdr);
Trond Myklebust493292d2011-07-13 15:58:28 -04001860}
1861
Benny Halevyd20581a2011-05-22 19:52:03 +03001862/*
Andy Adamson64419a92011-03-01 01:34:16 +00001863 * Call the appropriate parallel I/O subsystem read function.
1864 */
Trond Myklebust493292d2011-07-13 15:58:28 -04001865static enum pnfs_try_status
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001866pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
Trond Myklebust493292d2011-07-13 15:58:28 -04001867 const struct rpc_call_ops *call_ops,
1868 struct pnfs_layout_segment *lseg)
Andy Adamson64419a92011-03-01 01:34:16 +00001869{
Fred Isamancd841602012-04-20 14:47:44 -04001870 struct inode *inode = hdr->inode;
Andy Adamson64419a92011-03-01 01:34:16 +00001871 struct nfs_server *nfss = NFS_SERVER(inode);
1872 enum pnfs_try_status trypnfs;
1873
Fred Isamancd841602012-04-20 14:47:44 -04001874 hdr->mds_ops = call_ops;
Andy Adamson64419a92011-03-01 01:34:16 +00001875
1876 dprintk("%s: Reading ino:%lu %u@%llu\n",
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001877 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
Andy Adamson64419a92011-03-01 01:34:16 +00001878
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001879 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001880 if (trypnfs != PNFS_NOT_ATTEMPTED)
Andy Adamson64419a92011-03-01 01:34:16 +00001881 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
Andy Adamson64419a92011-03-01 01:34:16 +00001882 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1883 return trypnfs;
1884}
Andy Adamson863a3c62011-03-23 13:27:54 +00001885
Peng Taoceb11e12014-11-10 08:35:38 +08001886/* Resend all requests through pnfs. */
1887int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
1888{
1889 struct nfs_pageio_descriptor pgio;
1890
1891 nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
1892 return nfs_pageio_resend(&pgio, hdr);
1893}
1894EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
1895
Trond Myklebust493292d2011-07-13 15:58:28 -04001896static void
Weston Andros Adamson7f714722014-05-15 11:56:53 -04001897pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
Trond Myklebust493292d2011-07-13 15:58:28 -04001898{
Trond Myklebust493292d2011-07-13 15:58:28 -04001899 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1900 struct pnfs_layout_segment *lseg = desc->pg_lseg;
Weston Andros Adamson7f714722014-05-15 11:56:53 -04001901 enum pnfs_try_status trypnfs;
Peng Taoceb11e12014-11-10 08:35:38 +08001902 int err = 0;
Trond Myklebust493292d2011-07-13 15:58:28 -04001903
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001904 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
Peng Taoceb11e12014-11-10 08:35:38 +08001905 if (trypnfs == PNFS_TRY_AGAIN)
1906 err = pnfs_read_resend_pnfs(hdr);
1907 if (trypnfs == PNFS_NOT_ATTEMPTED || err)
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001908 pnfs_read_through_mds(desc, hdr);
Trond Myklebust493292d2011-07-13 15:58:28 -04001909}
1910
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001911static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1912{
Trond Myklebust9369a432012-09-18 20:57:08 -04001913 pnfs_put_lseg(hdr->lseg);
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04001914 nfs_pgio_header_free(hdr);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001915}
Bryan Schumaker89d77c82012-07-30 16:05:25 -04001916EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001917
Trond Myklebust493292d2011-07-13 15:58:28 -04001918int
1919pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1920{
Peng Tao48d635f2014-11-10 08:35:35 +08001921 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001922
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001923 struct nfs_pgio_header *hdr;
Trond Myklebust493292d2011-07-13 15:58:28 -04001924 int ret;
1925
Weston Andros Adamson1e7f3a42014-06-09 11:48:33 -04001926 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1927 if (!hdr) {
Weston Andros Adamsona7d42dd2014-09-19 10:55:07 -04001928 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04001929 return -ENOMEM;
Trond Myklebust493292d2011-07-13 15:58:28 -04001930 }
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001931 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
Trond Myklebust9369a432012-09-18 20:57:08 -04001932 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
Anna Schumakeref2c4882014-05-06 09:12:36 -04001933 ret = nfs_generic_pgio(desc, hdr);
Weston Andros Adamson180bb5e2014-09-10 15:48:01 -04001934 if (!ret)
Weston Andros Adamson7f714722014-05-15 11:56:53 -04001935 pnfs_do_read(desc, hdr);
Fred Isaman4db6e0b2012-04-20 14:47:46 -04001936 return ret;
Trond Myklebust493292d2011-07-13 15:58:28 -04001937}
1938EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1939
Trond Myklebust71244d92014-01-13 13:34:36 -05001940static void pnfs_clear_layoutcommitting(struct inode *inode)
1941{
1942 unsigned long *bitlock = &NFS_I(inode)->flags;
1943
1944 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
Peter Zijlstra4e857c52014-03-17 18:06:10 +01001945 smp_mb__after_atomic();
Trond Myklebust71244d92014-01-13 13:34:36 -05001946 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1947}
1948
Andy Adamson863a3c62011-03-23 13:27:54 +00001949/*
Peng Taoa9bae562011-07-30 20:52:33 -04001950 * There can be multiple RW segments.
Andy Adamson863a3c62011-03-23 13:27:54 +00001951 */
Peng Taoa9bae562011-07-30 20:52:33 -04001952static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
Andy Adamson863a3c62011-03-23 13:27:54 +00001953{
Peng Taoa9bae562011-07-30 20:52:33 -04001954 struct pnfs_layout_segment *lseg;
Andy Adamson863a3c62011-03-23 13:27:54 +00001955
Peng Taoa9bae562011-07-30 20:52:33 -04001956 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1957 if (lseg->pls_range.iomode == IOMODE_RW &&
Trond Myklebusta073dbf2013-03-20 12:34:32 -04001958 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
Peng Taoa9bae562011-07-30 20:52:33 -04001959 list_add(&lseg->pls_lc_list, listp);
1960 }
Andy Adamson863a3c62011-03-23 13:27:54 +00001961}
1962
Trond Myklebusta073dbf2013-03-20 12:34:32 -04001963static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1964{
1965 struct pnfs_layout_segment *lseg, *tmp;
Trond Myklebusta073dbf2013-03-20 12:34:32 -04001966
1967 /* Matched by references in pnfs_set_layoutcommit */
1968 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1969 list_del_init(&lseg->pls_lc_list);
1970 pnfs_put_lseg(lseg);
1971 }
1972
Trond Myklebust71244d92014-01-13 13:34:36 -05001973 pnfs_clear_layoutcommitting(inode);
Trond Myklebusta073dbf2013-03-20 12:34:32 -04001974}
1975
Peng Tao1b0ae062011-09-22 21:50:12 -04001976void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1977{
Trond Myklebustb9e028f2012-09-18 16:41:18 -04001978 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
Peng Tao1b0ae062011-09-22 21:50:12 -04001979}
1980EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1981
Andy Adamson863a3c62011-03-23 13:27:54 +00001982void
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001983pnfs_set_layoutcommit(struct nfs_pgio_header *hdr)
Andy Adamson863a3c62011-03-23 13:27:54 +00001984{
Fred Isamancd841602012-04-20 14:47:44 -04001985 struct inode *inode = hdr->inode;
1986 struct nfs_inode *nfsi = NFS_I(inode);
Weston Andros Adamsond45f60c2014-06-09 11:48:35 -04001987 loff_t end_pos = hdr->mds_offset + hdr->res.count;
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04001988 bool mark_as_dirty = false;
Andy Adamson863a3c62011-03-23 13:27:54 +00001989
Fred Isamancd841602012-04-20 14:47:44 -04001990 spin_lock(&inode->i_lock);
Andy Adamson863a3c62011-03-23 13:27:54 +00001991 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04001992 mark_as_dirty = true;
Andy Adamson863a3c62011-03-23 13:27:54 +00001993 dprintk("%s: Set layoutcommit for inode %lu ",
Fred Isamancd841602012-04-20 14:47:44 -04001994 __func__, inode->i_ino);
Andy Adamson863a3c62011-03-23 13:27:54 +00001995 }
Fred Isamancd841602012-04-20 14:47:44 -04001996 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
Peng Taoa9bae562011-07-30 20:52:33 -04001997 /* references matched in nfs4_layoutcommit_release */
Trond Myklebust9369a432012-09-18 20:57:08 -04001998 pnfs_get_lseg(hdr->lseg);
Peng Taoa9bae562011-07-30 20:52:33 -04001999 }
Peng Taoacff58802011-07-30 20:52:31 -04002000 if (end_pos > nfsi->layout->plh_lwb)
2001 nfsi->layout->plh_lwb = end_pos;
Fred Isamancd841602012-04-20 14:47:44 -04002002 spin_unlock(&inode->i_lock);
Peng Taoacff58802011-07-30 20:52:31 -04002003 dprintk("%s: lseg %p end_pos %llu\n",
Fred Isamancd841602012-04-20 14:47:44 -04002004 __func__, hdr->lseg, nfsi->layout->plh_lwb);
Weston Andros Adamson79a48a12011-04-13 10:53:51 -04002005
2006 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2007 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2008 if (mark_as_dirty)
Fred Isamancd841602012-04-20 14:47:44 -04002009 mark_inode_dirty_sync(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00002010}
2011EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2012
Peng Tao378520b2014-08-07 10:15:02 +08002013void pnfs_commit_set_layoutcommit(struct nfs_commit_data *data)
2014{
2015 struct inode *inode = data->inode;
2016 struct nfs_inode *nfsi = NFS_I(inode);
2017 bool mark_as_dirty = false;
2018
2019 spin_lock(&inode->i_lock);
2020 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2021 mark_as_dirty = true;
2022 dprintk("%s: Set layoutcommit for inode %lu ",
2023 __func__, inode->i_ino);
2024 }
2025 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &data->lseg->pls_flags)) {
2026 /* references matched in nfs4_layoutcommit_release */
2027 pnfs_get_lseg(data->lseg);
2028 }
2029 if (data->lwb > nfsi->layout->plh_lwb)
2030 nfsi->layout->plh_lwb = data->lwb;
2031 spin_unlock(&inode->i_lock);
2032 dprintk("%s: lseg %p end_pos %llu\n",
2033 __func__, data->lseg, nfsi->layout->plh_lwb);
2034
2035 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2036 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2037 if (mark_as_dirty)
2038 mark_inode_dirty_sync(inode);
2039}
2040EXPORT_SYMBOL_GPL(pnfs_commit_set_layoutcommit);
2041
Andy Adamsondb29c082011-07-30 20:52:38 -04002042void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2043{
2044 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2045
2046 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2047 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
Trond Myklebusta073dbf2013-03-20 12:34:32 -04002048 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
Andy Adamsondb29c082011-07-30 20:52:38 -04002049}
2050
Andy Adamsonde4b15c2011-03-12 02:58:09 -05002051/*
2052 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2053 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2054 * data to disk to allow the server to recover the data if it crashes.
2055 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2056 * is off, and a COMMIT is sent to a data server, or
2057 * if WRITEs to a data server return NFS_DATA_SYNC.
2058 */
Andy Adamson863a3c62011-03-23 13:27:54 +00002059int
Andy Adamsonef311532011-03-12 02:58:10 -05002060pnfs_layoutcommit_inode(struct inode *inode, bool sync)
Andy Adamson863a3c62011-03-23 13:27:54 +00002061{
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002062 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
Andy Adamson863a3c62011-03-23 13:27:54 +00002063 struct nfs4_layoutcommit_data *data;
2064 struct nfs_inode *nfsi = NFS_I(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00002065 loff_t end_pos;
Trond Myklebust71244d92014-01-13 13:34:36 -05002066 int status;
2067
2068 if (!pnfs_layoutcommit_outstanding(inode))
2069 return 0;
Andy Adamson863a3c62011-03-23 13:27:54 +00002070
2071 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2072
Trond Myklebust71244d92014-01-13 13:34:36 -05002073 status = -EAGAIN;
2074 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2075 if (!sync)
2076 goto out;
NeilBrown74316202014-07-07 15:16:04 +10002077 status = wait_on_bit_lock_action(&nfsi->flags,
Trond Myklebust71244d92014-01-13 13:34:36 -05002078 NFS_INO_LAYOUTCOMMITTING,
2079 nfs_wait_bit_killable,
2080 TASK_KILLABLE);
2081 if (status)
2082 goto out;
2083 }
Andy Adamsonde4b15c2011-03-12 02:58:09 -05002084
Trond Myklebust71244d92014-01-13 13:34:36 -05002085 status = -ENOMEM;
Andy Adamson863a3c62011-03-23 13:27:54 +00002086 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2087 data = kzalloc(sizeof(*data), GFP_NOFS);
Trond Myklebust71244d92014-01-13 13:34:36 -05002088 if (!data)
2089 goto clear_layoutcommitting;
Andy Adamson863a3c62011-03-23 13:27:54 +00002090
Trond Myklebust71244d92014-01-13 13:34:36 -05002091 status = 0;
2092 spin_lock(&inode->i_lock);
2093 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2094 goto out_unlock;
Peng Tao92407e72011-10-23 20:21:17 -07002095
Peng Taoa9bae562011-07-30 20:52:33 -04002096 INIT_LIST_HEAD(&data->lseg_list);
Peng Taoa9bae562011-07-30 20:52:33 -04002097 pnfs_list_write_lseg(inode, &data->lseg_list);
Andy Adamson863a3c62011-03-23 13:27:54 +00002098
Peng Taoacff58802011-07-30 20:52:31 -04002099 end_pos = nfsi->layout->plh_lwb;
Peng Taoacff58802011-07-30 20:52:31 -04002100 nfsi->layout->plh_lwb = 0;
Andy Adamson863a3c62011-03-23 13:27:54 +00002101
Trond Myklebustf597c532012-03-04 18:13:56 -05002102 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
Andy Adamson863a3c62011-03-23 13:27:54 +00002103 spin_unlock(&inode->i_lock);
2104
2105 data->args.inode = inode;
Peng Tao9fa40752011-07-30 20:52:32 -04002106 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
Andy Adamson863a3c62011-03-23 13:27:54 +00002107 nfs_fattr_init(&data->fattr);
2108 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2109 data->res.fattr = &data->fattr;
2110 data->args.lastbytewritten = end_pos - 1;
2111 data->res.server = NFS_SERVER(inode);
2112
Christoph Hellwig5f919c92014-08-21 11:09:25 -05002113 if (ld->prepare_layoutcommit) {
2114 status = ld->prepare_layoutcommit(&data->args);
2115 if (status) {
2116 spin_lock(&inode->i_lock);
2117 if (end_pos < nfsi->layout->plh_lwb)
2118 nfsi->layout->plh_lwb = end_pos;
2119 spin_unlock(&inode->i_lock);
2120 put_rpccred(data->cred);
2121 set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2122 goto clear_layoutcommitting;
2123 }
2124 }
2125
2126
Andy Adamson863a3c62011-03-23 13:27:54 +00002127 status = nfs4_proc_layoutcommit(data, sync);
2128out:
Peng Tao92407e72011-10-23 20:21:17 -07002129 if (status)
2130 mark_inode_dirty_sync(inode);
Andy Adamson863a3c62011-03-23 13:27:54 +00002131 dprintk("<-- %s status %d\n", __func__, status);
2132 return status;
Trond Myklebust71244d92014-01-13 13:34:36 -05002133out_unlock:
2134 spin_unlock(&inode->i_lock);
Peng Tao92407e72011-10-23 20:21:17 -07002135 kfree(data);
Trond Myklebust71244d92014-01-13 13:34:36 -05002136clear_layoutcommitting:
2137 pnfs_clear_layoutcommitting(inode);
Peng Tao92407e72011-10-23 20:21:17 -07002138 goto out;
Andy Adamson863a3c62011-03-23 13:27:54 +00002139}
Peng Tao72cff442014-08-07 10:12:38 +08002140EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
Andy Adamson82be4172012-05-23 05:02:35 -04002141
2142struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2143{
2144 struct nfs4_threshold *thp;
2145
2146 thp = kzalloc(sizeof(*thp), GFP_NOFS);
2147 if (!thp) {
2148 dprintk("%s mdsthreshold allocation failed\n", __func__);
2149 return NULL;
2150 }
2151 return thp;
2152}