blob: 21e921147d21b7b42253deb2621e208ba59c18af [file] [log] [blame]
Dean Hildebrand7ab672c2010-10-20 00:18:00 -04001/*
2 * Module for the pnfs nfs4 file layout driver.
3 * Defines all I/O and Policy interface operations, plus code
4 * to register itself with the pNFS client.
5 *
6 * Copyright (c) 2002
7 * The Regents of the University of Michigan
8 * All Rights Reserved
9 *
10 * Dean Hildebrand <dhildebz@umich.edu>
11 *
12 * Permission is granted to use, copy, create derivative works, and
13 * redistribute this software and such derivative works for any purpose,
14 * so long as the name of the University of Michigan is not used in
15 * any advertising or publicity pertaining to the use or distribution
16 * of this software without specific, written prior authorization. If
17 * the above copyright notice or any other identification of the
18 * University of Michigan is included in any copy of any portion of
19 * this software, then the disclaimer below must also be included.
20 *
21 * This software is provided as is, without representation or warranty
22 * of any kind either express or implied, including without limitation
23 * the implied warranties of merchantability, fitness for a particular
24 * purpose, or noninfringement. The Regents of the University of
25 * Michigan shall not be liable for any damages, including special,
26 * indirect, incidental, or consequential damages, with respect to any
27 * claim arising out of or in connection with the use of the software,
28 * even if it has been or is hereafter advised of the possibility of
29 * such damages.
30 */
31
32#include <linux/nfs_fs.h>
Benny Halevy19345cb2011-06-19 18:33:46 -040033#include <linux/nfs_page.h>
Paul Gortmaker143cb492011-07-01 14:23:34 -040034#include <linux/module.h>
Andy Adamson16b374c2010-10-20 00:18:04 -040035
Weston Andros Adamson0a702192012-02-17 13:15:24 -050036#include <linux/sunrpc/metrics.h>
37
Trond Myklebust76e697b2012-11-26 14:20:49 -050038#include "nfs4session.h"
Andy Adamson16b374c2010-10-20 00:18:04 -040039#include "internal.h"
Andy Adamson9cb81962012-03-07 10:49:41 -050040#include "delegation.h"
Andy Adamson16b374c2010-10-20 00:18:04 -040041#include "nfs4filelayout.h"
Trond Myklebustcc668ab2013-08-14 15:31:28 -040042#include "nfs4trace.h"
Dean Hildebrand7ab672c2010-10-20 00:18:00 -040043
44#define NFSDBG_FACILITY NFSDBG_PNFS_LD
45
46MODULE_LICENSE("GPL");
47MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
48MODULE_DESCRIPTION("The NFSv4 file layout driver");
49
Andy Adamsoncbdabc72011-03-01 01:34:20 +000050#define FILELAYOUT_POLL_RETRY_MAX (15*HZ)
51
Fred Isamancfe7f412011-03-01 01:34:18 +000052static loff_t
53filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
54 loff_t offset)
55{
56 u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
Chris Metcalf3476f112011-08-11 13:54:28 -070057 u64 stripe_no;
58 u32 rem;
Fred Isamancfe7f412011-03-01 01:34:18 +000059
60 offset -= flseg->pattern_offset;
Chris Metcalf3476f112011-08-11 13:54:28 -070061 stripe_no = div_u64(offset, stripe_width);
62 div_u64_rem(offset, flseg->stripe_unit, &rem);
Fred Isamancfe7f412011-03-01 01:34:18 +000063
Chris Metcalf3476f112011-08-11 13:54:28 -070064 return stripe_no * flseg->stripe_unit + rem;
Fred Isamancfe7f412011-03-01 01:34:18 +000065}
66
67/* This function is used by the layout driver to calculate the
68 * offset of the file on the dserver based on whether the
69 * layout type is STRIPE_DENSE or STRIPE_SPARSE
70 */
71static loff_t
72filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
73{
74 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
75
76 switch (flseg->stripe_type) {
77 case STRIPE_SPARSE:
78 return offset;
79
80 case STRIPE_DENSE:
81 return filelayout_get_dense_offset(flseg, offset);
82 }
83
84 BUG();
85}
86
Anna Schumaker9c7e1b32014-05-06 09:12:26 -040087static void filelayout_reset_write(struct nfs_pgio_data *data)
Andy Adamsone7dd79af2012-04-27 17:53:46 -040088{
89 struct nfs_pgio_header *hdr = data->header;
Andy Adamsone7dd79af2012-04-27 17:53:46 -040090 struct rpc_task *task = &data->task;
91
92 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
93 dprintk("%s Reset task %5u for i/o through MDS "
Niels de Vos1e8968c2013-12-17 18:20:16 +010094 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
Andy Adamsone7dd79af2012-04-27 17:53:46 -040095 data->task.tk_pid,
Bryan Schumaker497826a2012-05-22 10:10:03 -040096 hdr->inode->i_sb->s_id,
Niels de Vos1e8968c2013-12-17 18:20:16 +010097 (unsigned long long)NFS_FILEID(hdr->inode),
Andy Adamsone7dd79af2012-04-27 17:53:46 -040098 data->args.count,
99 (unsigned long long)data->args.offset);
100
101 task->tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
102 &hdr->pages,
Benny Halevy78f33272013-02-24 09:55:57 -0500103 hdr->completion_ops,
104 hdr->dreq);
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400105 }
106}
107
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400108static void filelayout_reset_read(struct nfs_pgio_data *data)
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400109{
110 struct nfs_pgio_header *hdr = data->header;
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400111 struct rpc_task *task = &data->task;
112
113 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
114 dprintk("%s Reset task %5u for i/o through MDS "
Niels de Vos1e8968c2013-12-17 18:20:16 +0100115 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400116 data->task.tk_pid,
Bryan Schumaker497826a2012-05-22 10:10:03 -0400117 hdr->inode->i_sb->s_id,
Niels de Vos1e8968c2013-12-17 18:20:16 +0100118 (unsigned long long)NFS_FILEID(hdr->inode),
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400119 data->args.count,
120 (unsigned long long)data->args.offset);
121
122 task->tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
123 &hdr->pages,
Benny Halevy78f33272013-02-24 09:55:57 -0500124 hdr->completion_ops,
125 hdr->dreq);
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400126 }
127}
128
Trond Myklebustd527e5c2012-10-11 13:43:38 -0400129static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo)
130{
131 if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
132 return;
Trond Myklebustd527e5c2012-10-11 13:43:38 -0400133 pnfs_return_layout(inode);
134}
135
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000136static int filelayout_async_handle_error(struct rpc_task *task,
137 struct nfs4_state *state,
138 struct nfs_client *clp,
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400139 struct pnfs_layout_segment *lseg)
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000140{
Trond Myklebustd527e5c2012-10-11 13:43:38 -0400141 struct pnfs_layout_hdr *lo = lseg->pls_layout;
142 struct inode *inode = lo->plh_inode;
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400143 struct nfs_server *mds_server = NFS_SERVER(inode);
144 struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
Andy Adamson9cb81962012-03-07 10:49:41 -0500145 struct nfs_client *mds_client = mds_server->nfs_client;
Andy Adamson671fb892012-04-27 17:53:49 -0400146 struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
Andy Adamson9cb81962012-03-07 10:49:41 -0500147
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000148 if (task->tk_status >= 0)
149 return 0;
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000150
151 switch (task->tk_status) {
Andy Adamson9cb81962012-03-07 10:49:41 -0500152 /* MDS state errors */
153 case -NFS4ERR_DELEG_REVOKED:
154 case -NFS4ERR_ADMIN_REVOKED:
155 case -NFS4ERR_BAD_STATEID:
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400156 if (state == NULL)
157 break;
Andy Adamson2dc31752012-03-08 11:03:53 -0500158 nfs_remove_bad_delegation(state->inode);
Andy Adamson9cb81962012-03-07 10:49:41 -0500159 case -NFS4ERR_OPENMODE:
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400160 if (state == NULL)
161 break;
Trond Myklebust5d422302013-03-14 16:57:48 -0400162 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
163 goto out_bad_stateid;
Andy Adamson9cb81962012-03-07 10:49:41 -0500164 goto wait_on_recovery;
165 case -NFS4ERR_EXPIRED:
Trond Myklebust5d422302013-03-14 16:57:48 -0400166 if (state != NULL) {
167 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
168 goto out_bad_stateid;
169 }
Andy Adamson9cb81962012-03-07 10:49:41 -0500170 nfs4_schedule_lease_recovery(mds_client);
171 goto wait_on_recovery;
172 /* DS session errors */
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000173 case -NFS4ERR_BADSESSION:
174 case -NFS4ERR_BADSLOT:
175 case -NFS4ERR_BAD_HIGH_SLOT:
176 case -NFS4ERR_DEADSESSION:
177 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
178 case -NFS4ERR_SEQ_FALSE_RETRY:
179 case -NFS4ERR_SEQ_MISORDERED:
180 dprintk("%s ERROR %d, Reset session. Exchangeid "
181 "flags 0x%x\n", __func__, task->tk_status,
182 clp->cl_exchange_flags);
Trond Myklebust9f594792012-05-27 13:02:53 -0400183 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000184 break;
185 case -NFS4ERR_DELAY:
186 case -NFS4ERR_GRACE:
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000187 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
188 break;
Andy Adamsona8a4ae32011-05-03 13:43:03 -0400189 case -NFS4ERR_RETRY_UNCACHED_REP:
190 break;
Andy Adamson041245c2012-04-27 17:53:53 -0400191 /* Invalidate Layout errors */
192 case -NFS4ERR_PNFS_NO_LAYOUT:
193 case -ESTALE: /* mapped NFS4ERR_STALE */
194 case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */
195 case -EISDIR: /* mapped NFS4ERR_ISDIR */
196 case -NFS4ERR_FHEXPIRED:
197 case -NFS4ERR_WRONG_TYPE:
198 dprintk("%s Invalid layout error %d\n", __func__,
199 task->tk_status);
200 /*
201 * Destroy layout so new i/o will get a new layout.
202 * Layout will not be destroyed until all current lseg
203 * references are put. Mark layout as invalid to resend failed
204 * i/o and all i/o waiting on the slot table to the MDS until
205 * layout is destroyed and a new valid layout is obtained.
206 */
Andy Adamsond42e7872012-05-22 08:09:28 -0400207 pnfs_destroy_layout(NFS_I(inode));
Andy Adamson041245c2012-04-27 17:53:53 -0400208 rpc_wake_up(&tbl->slot_tbl_waitq);
209 goto reset;
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400210 /* RPC connection errors */
211 case -ECONNREFUSED:
212 case -EHOSTDOWN:
213 case -EHOSTUNREACH:
214 case -ENETUNREACH:
215 case -EIO:
216 case -ETIMEDOUT:
217 case -EPIPE:
218 dprintk("%s DS connection error %d\n", __func__,
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000219 task->tk_status);
Trond Myklebust1dfed272012-09-18 19:51:12 -0400220 nfs4_mark_deviceid_unavailable(devid);
Trond Myklebustd527e5c2012-10-11 13:43:38 -0400221 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
Andy Adamson671fb892012-04-27 17:53:49 -0400222 rpc_wake_up(&tbl->slot_tbl_waitq);
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400223 /* fall through */
224 default:
Andy Adamson041245c2012-04-27 17:53:53 -0400225reset:
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400226 dprintk("%s Retry through MDS. Error %d\n", __func__,
227 task->tk_status);
228 return -NFS4ERR_RESET_TO_MDS;
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000229 }
Andy Adamson9cb81962012-03-07 10:49:41 -0500230out:
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000231 task->tk_status = 0;
232 return -EAGAIN;
Trond Myklebust5d422302013-03-14 16:57:48 -0400233out_bad_stateid:
234 task->tk_status = -EIO;
235 return 0;
Andy Adamson9cb81962012-03-07 10:49:41 -0500236wait_on_recovery:
237 rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
238 if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
239 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
240 goto out;
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000241}
242
243/* NFS_PROTO call done callback routines */
244
245static int filelayout_read_done_cb(struct rpc_task *task,
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400246 struct nfs_pgio_data *data)
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000247{
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400248 struct nfs_pgio_header *hdr = data->header;
249 int err;
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000250
Trond Myklebustcc668ab2013-08-14 15:31:28 -0400251 trace_nfs4_pnfs_read(data, task->tk_status);
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400252 err = filelayout_async_handle_error(task, data->args.context->state,
253 data->ds_clp, hdr->lseg);
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000254
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400255 switch (err) {
256 case -NFS4ERR_RESET_TO_MDS:
257 filelayout_reset_read(data);
258 return task->tk_status;
259 case -EAGAIN:
Trond Myklebustd00c5d42011-10-19 12:17:29 -0700260 rpc_restart_call_prepare(task);
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000261 return -EAGAIN;
262 }
263
264 return 0;
265}
266
Andy Adamson16b374c2010-10-20 00:18:04 -0400267/*
Andy Adamson863a3c62011-03-23 13:27:54 +0000268 * We reference the rpc_cred of the first WRITE that triggers the need for
269 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
270 * rfc5661 is not clear about which credential should be used.
271 */
272static void
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400273filelayout_set_layoutcommit(struct nfs_pgio_data *wdata)
Andy Adamson863a3c62011-03-23 13:27:54 +0000274{
Fred Isamancd841602012-04-20 14:47:44 -0400275 struct nfs_pgio_header *hdr = wdata->header;
276
277 if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
Andy Adamson863a3c62011-03-23 13:27:54 +0000278 wdata->res.verf->committed == NFS_FILE_SYNC)
279 return;
280
281 pnfs_set_layoutcommit(wdata);
Fred Isamancd841602012-04-20 14:47:44 -0400282 dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
283 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
Andy Adamson863a3c62011-03-23 13:27:54 +0000284}
285
Trond Myklebust1dfed272012-09-18 19:51:12 -0400286bool
287filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
288{
289 return filelayout_test_devid_invalid(node) ||
290 nfs4_test_deviceid_unavailable(node);
291}
292
293static bool
294filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
295{
296 struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
297
Trond Myklebust8006bfb2012-09-21 14:48:04 -0400298 return filelayout_test_devid_unavailable(node);
Trond Myklebust1dfed272012-09-18 19:51:12 -0400299}
300
Andy Adamson863a3c62011-03-23 13:27:54 +0000301/*
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000302 * Call ops for the async read/write cases
303 * In the case of dense layouts, the offset needs to be reset to its
304 * original value.
305 */
306static void filelayout_read_prepare(struct rpc_task *task, void *data)
307{
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400308 struct nfs_pgio_data *rdata = data;
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000309
Trond Myklebustc58c8442013-03-18 19:45:14 -0400310 if (unlikely(test_bit(NFS_CONTEXT_BAD, &rdata->args.context->flags))) {
311 rpc_exit(task, -EIO);
312 return;
313 }
Andy Adamson041245c2012-04-27 17:53:53 -0400314 if (filelayout_reset_to_mds(rdata->header->lseg)) {
Andy Adamson0ad2f372012-04-27 17:53:48 -0400315 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
316 filelayout_reset_read(rdata);
317 rpc_exit(task, 0);
318 return;
319 }
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400320 rdata->pgio_done_cb = filelayout_read_done_cb;
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000321
Trond Myklebust9b206142013-03-17 15:52:00 -0400322 if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
Trond Myklebustd9afbd12012-10-22 20:28:44 -0400323 &rdata->args.seq_args,
324 &rdata->res.seq_res,
Trond Myklebust9b206142013-03-17 15:52:00 -0400325 task))
326 return;
Andy Adamson869a9d32014-03-04 12:31:09 -0500327 if (nfs4_set_rw_stateid(&rdata->args.stateid, rdata->args.context,
328 rdata->args.lock_context, FMODE_READ) == -EIO)
329 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000330}
331
332static void filelayout_read_call_done(struct rpc_task *task, void *data)
333{
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400334 struct nfs_pgio_data *rdata = data;
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000335
336 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
337
Andy Adamsonbd4aeff2012-05-22 08:09:27 -0400338 if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) &&
Andy Adamsonf9c96fc2014-01-29 11:34:38 -0500339 task->tk_status == 0) {
Trond Myklebust905e7da2014-01-29 12:26:57 -0500340 nfs41_sequence_done(task, &rdata->res.seq_res);
Andy Adamson0ad2f372012-04-27 17:53:48 -0400341 return;
Andy Adamsonf9c96fc2014-01-29 11:34:38 -0500342 }
Andy Adamson0ad2f372012-04-27 17:53:48 -0400343
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000344 /* Note this may cause RPC to be resent */
Fred Isamancd841602012-04-20 14:47:44 -0400345 rdata->header->mds_ops->rpc_call_done(task, data);
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000346}
347
Weston Andros Adamson0a702192012-02-17 13:15:24 -0500348static void filelayout_read_count_stats(struct rpc_task *task, void *data)
349{
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400350 struct nfs_pgio_data *rdata = data;
Weston Andros Adamson0a702192012-02-17 13:15:24 -0500351
Fred Isamancd841602012-04-20 14:47:44 -0400352 rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
Weston Andros Adamson0a702192012-02-17 13:15:24 -0500353}
354
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000355static void filelayout_read_release(void *data)
356{
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400357 struct nfs_pgio_data *rdata = data;
Trond Myklebustd527e5c2012-10-11 13:43:38 -0400358 struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout;
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000359
Trond Myklebustd527e5c2012-10-11 13:43:38 -0400360 filelayout_fenceme(lo->plh_inode, lo);
Andy Adamson996074c2012-05-22 08:09:26 -0400361 nfs_put_client(rdata->ds_clp);
Fred Isamancd841602012-04-20 14:47:44 -0400362 rdata->header->mds_ops->rpc_release(data);
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000363}
364
Fred Isamana69aef12011-03-03 15:13:47 +0000365static int filelayout_write_done_cb(struct rpc_task *task,
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400366 struct nfs_pgio_data *data)
Fred Isamana69aef12011-03-03 15:13:47 +0000367{
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400368 struct nfs_pgio_header *hdr = data->header;
369 int err;
Fred Isamana69aef12011-03-03 15:13:47 +0000370
Trond Myklebustcc668ab2013-08-14 15:31:28 -0400371 trace_nfs4_pnfs_write(data, task->tk_status);
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400372 err = filelayout_async_handle_error(task, data->args.context->state,
373 data->ds_clp, hdr->lseg);
374
375 switch (err) {
376 case -NFS4ERR_RESET_TO_MDS:
377 filelayout_reset_write(data);
378 return task->tk_status;
379 case -EAGAIN:
Trond Myklebustd00c5d42011-10-19 12:17:29 -0700380 rpc_restart_call_prepare(task);
Fred Isamana69aef12011-03-03 15:13:47 +0000381 return -EAGAIN;
382 }
383
Andy Adamson863a3c62011-03-23 13:27:54 +0000384 filelayout_set_layoutcommit(data);
Fred Isamana69aef12011-03-03 15:13:47 +0000385 return 0;
386}
387
Fred Isamane0c2b382011-03-23 13:27:53 +0000388/* Fake up some data that will cause nfs_commit_release to retry the writes. */
Fred Isaman0b7c0152012-04-20 14:47:39 -0400389static void prepare_to_resend_writes(struct nfs_commit_data *data)
Fred Isamane0c2b382011-03-23 13:27:53 +0000390{
391 struct nfs_page *first = nfs_list_entry(data->pages.next);
392
393 data->task.tk_status = 0;
Trond Myklebust2f2c63b2012-06-08 11:56:09 -0400394 memcpy(&data->verf.verifier, &first->wb_verf,
395 sizeof(data->verf.verifier));
396 data->verf.verifier.data[0]++; /* ensure verifier mismatch */
Fred Isamane0c2b382011-03-23 13:27:53 +0000397}
398
399static int filelayout_commit_done_cb(struct rpc_task *task,
Fred Isaman0b7c0152012-04-20 14:47:39 -0400400 struct nfs_commit_data *data)
Fred Isamane0c2b382011-03-23 13:27:53 +0000401{
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400402 int err;
Fred Isamane0c2b382011-03-23 13:27:53 +0000403
Trond Myklebustcc668ab2013-08-14 15:31:28 -0400404 trace_nfs4_pnfs_commit_ds(data, task->tk_status);
Andy Adamsone7dd79af2012-04-27 17:53:46 -0400405 err = filelayout_async_handle_error(task, NULL, data->ds_clp,
406 data->lseg);
407
408 switch (err) {
409 case -NFS4ERR_RESET_TO_MDS:
410 prepare_to_resend_writes(data);
411 return -EAGAIN;
412 case -EAGAIN:
413 rpc_restart_call_prepare(task);
Fred Isamane0c2b382011-03-23 13:27:53 +0000414 return -EAGAIN;
415 }
416
417 return 0;
418}
419
Fred Isamana69aef12011-03-03 15:13:47 +0000420static void filelayout_write_prepare(struct rpc_task *task, void *data)
421{
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400422 struct nfs_pgio_data *wdata = data;
Fred Isamana69aef12011-03-03 15:13:47 +0000423
Trond Myklebustc58c8442013-03-18 19:45:14 -0400424 if (unlikely(test_bit(NFS_CONTEXT_BAD, &wdata->args.context->flags))) {
425 rpc_exit(task, -EIO);
426 return;
427 }
Andy Adamson041245c2012-04-27 17:53:53 -0400428 if (filelayout_reset_to_mds(wdata->header->lseg)) {
Andy Adamson0ad2f372012-04-27 17:53:48 -0400429 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
430 filelayout_reset_write(wdata);
431 rpc_exit(task, 0);
432 return;
433 }
Trond Myklebust9b206142013-03-17 15:52:00 -0400434 if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
Trond Myklebustd9afbd12012-10-22 20:28:44 -0400435 &wdata->args.seq_args,
436 &wdata->res.seq_res,
Trond Myklebust9b206142013-03-17 15:52:00 -0400437 task))
438 return;
Andy Adamson869a9d32014-03-04 12:31:09 -0500439 if (nfs4_set_rw_stateid(&wdata->args.stateid, wdata->args.context,
440 wdata->args.lock_context, FMODE_WRITE) == -EIO)
441 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
Fred Isamana69aef12011-03-03 15:13:47 +0000442}
443
444static void filelayout_write_call_done(struct rpc_task *task, void *data)
445{
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400446 struct nfs_pgio_data *wdata = data;
Fred Isamana69aef12011-03-03 15:13:47 +0000447
Andy Adamsonbd4aeff2012-05-22 08:09:27 -0400448 if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) &&
Andy Adamsonf9c96fc2014-01-29 11:34:38 -0500449 task->tk_status == 0) {
Trond Myklebust905e7da2014-01-29 12:26:57 -0500450 nfs41_sequence_done(task, &wdata->res.seq_res);
Andy Adamson0ad2f372012-04-27 17:53:48 -0400451 return;
Andy Adamsonf9c96fc2014-01-29 11:34:38 -0500452 }
Andy Adamson0ad2f372012-04-27 17:53:48 -0400453
Fred Isamana69aef12011-03-03 15:13:47 +0000454 /* Note this may cause RPC to be resent */
Fred Isamancd841602012-04-20 14:47:44 -0400455 wdata->header->mds_ops->rpc_call_done(task, data);
Fred Isamana69aef12011-03-03 15:13:47 +0000456}
457
Weston Andros Adamson0a702192012-02-17 13:15:24 -0500458static void filelayout_write_count_stats(struct rpc_task *task, void *data)
459{
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400460 struct nfs_pgio_data *wdata = data;
Weston Andros Adamson0a702192012-02-17 13:15:24 -0500461
Fred Isamancd841602012-04-20 14:47:44 -0400462 rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
Weston Andros Adamson0a702192012-02-17 13:15:24 -0500463}
464
Fred Isamana69aef12011-03-03 15:13:47 +0000465static void filelayout_write_release(void *data)
466{
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400467 struct nfs_pgio_data *wdata = data;
Trond Myklebustd527e5c2012-10-11 13:43:38 -0400468 struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout;
Fred Isamana69aef12011-03-03 15:13:47 +0000469
Trond Myklebustd527e5c2012-10-11 13:43:38 -0400470 filelayout_fenceme(lo->plh_inode, lo);
Andy Adamson996074c2012-05-22 08:09:26 -0400471 nfs_put_client(wdata->ds_clp);
Fred Isamancd841602012-04-20 14:47:44 -0400472 wdata->header->mds_ops->rpc_release(data);
Fred Isamana69aef12011-03-03 15:13:47 +0000473}
474
Fred Isaman0b7c0152012-04-20 14:47:39 -0400475static void filelayout_commit_prepare(struct rpc_task *task, void *data)
Fred Isamane0c2b382011-03-23 13:27:53 +0000476{
Fred Isaman0b7c0152012-04-20 14:47:39 -0400477 struct nfs_commit_data *wdata = data;
Fred Isamane0c2b382011-03-23 13:27:53 +0000478
Trond Myklebustd9afbd12012-10-22 20:28:44 -0400479 nfs41_setup_sequence(wdata->ds_clp->cl_session,
480 &wdata->args.seq_args,
481 &wdata->res.seq_res,
482 task);
Fred Isaman0b7c0152012-04-20 14:47:39 -0400483}
484
485static void filelayout_write_commit_done(struct rpc_task *task, void *data)
486{
487 struct nfs_commit_data *wdata = data;
488
489 /* Note this may cause RPC to be resent */
490 wdata->mds_ops->rpc_call_done(task, data);
491}
492
493static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
494{
495 struct nfs_commit_data *cdata = data;
496
497 rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
498}
499
500static void filelayout_commit_release(void *calldata)
501{
502 struct nfs_commit_data *data = calldata;
503
Fred Isamanf453a542012-04-20 14:47:54 -0400504 data->completion_ops->completion(data);
Trond Myklebust9369a432012-09-18 20:57:08 -0400505 pnfs_put_lseg(data->lseg);
Andy Adamson3a7936c2012-04-27 17:53:51 -0400506 nfs_put_client(data->ds_clp);
Fred Isaman0b7c0152012-04-20 14:47:39 -0400507 nfs_commitdata_release(data);
Fred Isamane0c2b382011-03-23 13:27:53 +0000508}
509
Trond Myklebust17280172012-03-11 13:11:00 -0400510static const struct rpc_call_ops filelayout_read_call_ops = {
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000511 .rpc_call_prepare = filelayout_read_prepare,
512 .rpc_call_done = filelayout_read_call_done,
Weston Andros Adamson0a702192012-02-17 13:15:24 -0500513 .rpc_count_stats = filelayout_read_count_stats,
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000514 .rpc_release = filelayout_read_release,
515};
516
Trond Myklebust17280172012-03-11 13:11:00 -0400517static const struct rpc_call_ops filelayout_write_call_ops = {
Fred Isamana69aef12011-03-03 15:13:47 +0000518 .rpc_call_prepare = filelayout_write_prepare,
519 .rpc_call_done = filelayout_write_call_done,
Weston Andros Adamson0a702192012-02-17 13:15:24 -0500520 .rpc_count_stats = filelayout_write_count_stats,
Fred Isamana69aef12011-03-03 15:13:47 +0000521 .rpc_release = filelayout_write_release,
522};
523
Trond Myklebust17280172012-03-11 13:11:00 -0400524static const struct rpc_call_ops filelayout_commit_call_ops = {
Fred Isaman0b7c0152012-04-20 14:47:39 -0400525 .rpc_call_prepare = filelayout_commit_prepare,
526 .rpc_call_done = filelayout_write_commit_done,
527 .rpc_count_stats = filelayout_commit_count_stats,
Fred Isamane0c2b382011-03-23 13:27:53 +0000528 .rpc_release = filelayout_commit_release,
529};
530
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000531static enum pnfs_try_status
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400532filelayout_read_pagelist(struct nfs_pgio_data *data)
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000533{
Fred Isamancd841602012-04-20 14:47:44 -0400534 struct nfs_pgio_header *hdr = data->header;
535 struct pnfs_layout_segment *lseg = hdr->lseg;
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000536 struct nfs4_pnfs_ds *ds;
Andy Adamson0e201622013-09-06 14:14:00 -0400537 struct rpc_clnt *ds_clnt;
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000538 loff_t offset = data->args.offset;
539 u32 j, idx;
540 struct nfs_fh *fh;
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000541
542 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
Fred Isamancd841602012-04-20 14:47:44 -0400543 __func__, hdr->inode->i_ino,
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000544 data->args.pgbase, (size_t)data->args.count, offset);
545
546 /* Retrieve the correct rpc_client for the byte range */
547 j = nfs4_fl_calc_j_index(lseg, offset);
548 idx = nfs4_fl_calc_ds_index(lseg, j);
549 ds = nfs4_fl_prepare_ds(lseg, idx);
Andy Adamson90fecfc2012-04-27 17:53:43 -0400550 if (!ds)
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000551 return PNFS_NOT_ATTEMPTED;
Andy Adamson0e201622013-09-06 14:14:00 -0400552
553 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
554 if (IS_ERR(ds_clnt))
555 return PNFS_NOT_ATTEMPTED;
556
Andy Adamson3a7936c2012-04-27 17:53:51 -0400557 dprintk("%s USE DS: %s cl_count %d\n", __func__,
558 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000559
560 /* No multipath support. Use first DS */
Andy Adamson3a7936c2012-04-27 17:53:51 -0400561 atomic_inc(&ds->ds_clp->cl_count);
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000562 data->ds_clp = ds->ds_clp;
Weston Andros Adamson5002c582014-05-15 11:56:54 -0400563 data->ds_idx = idx;
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000564 fh = nfs4_fl_select_ds_fh(lseg, j);
565 if (fh)
566 data->args.fh = fh;
567
568 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
569 data->mds_offset = offset;
570
571 /* Perform an asynchronous read to ds */
Anna Schumaker1ed26f32014-05-06 09:12:37 -0400572 nfs_initiate_pgio(ds_clnt, data,
573 &filelayout_read_call_ops, 0, RPC_TASK_SOFTCONN);
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000574 return PNFS_ATTEMPTED;
575}
576
Fred Isamana69aef12011-03-03 15:13:47 +0000577/* Perform async writes. */
Andy Adamson0382b742011-03-03 15:13:45 +0000578static enum pnfs_try_status
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400579filelayout_write_pagelist(struct nfs_pgio_data *data, int sync)
Andy Adamson0382b742011-03-03 15:13:45 +0000580{
Fred Isamancd841602012-04-20 14:47:44 -0400581 struct nfs_pgio_header *hdr = data->header;
582 struct pnfs_layout_segment *lseg = hdr->lseg;
Fred Isamana69aef12011-03-03 15:13:47 +0000583 struct nfs4_pnfs_ds *ds;
Andy Adamson0e201622013-09-06 14:14:00 -0400584 struct rpc_clnt *ds_clnt;
Fred Isamana69aef12011-03-03 15:13:47 +0000585 loff_t offset = data->args.offset;
586 u32 j, idx;
587 struct nfs_fh *fh;
Fred Isamana69aef12011-03-03 15:13:47 +0000588
589 /* Retrieve the correct rpc_client for the byte range */
590 j = nfs4_fl_calc_j_index(lseg, offset);
591 idx = nfs4_fl_calc_ds_index(lseg, j);
592 ds = nfs4_fl_prepare_ds(lseg, idx);
Andy Adamson90fecfc2012-04-27 17:53:43 -0400593 if (!ds)
Fred Isamana69aef12011-03-03 15:13:47 +0000594 return PNFS_NOT_ATTEMPTED;
Andy Adamson0e201622013-09-06 14:14:00 -0400595
596 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
597 if (IS_ERR(ds_clnt))
598 return PNFS_NOT_ATTEMPTED;
599
Andy Adamson3a7936c2012-04-27 17:53:51 -0400600 dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
601 __func__, hdr->inode->i_ino, sync, (size_t) data->args.count,
602 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
Fred Isamana69aef12011-03-03 15:13:47 +0000603
Anna Schumaker9c7e1b32014-05-06 09:12:26 -0400604 data->pgio_done_cb = filelayout_write_done_cb;
Andy Adamson3a7936c2012-04-27 17:53:51 -0400605 atomic_inc(&ds->ds_clp->cl_count);
Fred Isamana69aef12011-03-03 15:13:47 +0000606 data->ds_clp = ds->ds_clp;
Weston Andros Adamson5002c582014-05-15 11:56:54 -0400607 data->ds_idx = idx;
Fred Isamana69aef12011-03-03 15:13:47 +0000608 fh = nfs4_fl_select_ds_fh(lseg, j);
609 if (fh)
610 data->args.fh = fh;
611 /*
612 * Get the file offset on the dserver. Set the write offset to
613 * this offset and save the original offset.
614 */
615 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
Fred Isamana69aef12011-03-03 15:13:47 +0000616
617 /* Perform an asynchronous write */
Anna Schumaker1ed26f32014-05-06 09:12:37 -0400618 nfs_initiate_pgio(ds_clnt, data,
Andy Adamson9f0ec1762012-04-27 17:53:44 -0400619 &filelayout_write_call_ops, sync,
620 RPC_TASK_SOFTCONN);
Fred Isamana69aef12011-03-03 15:13:47 +0000621 return PNFS_ATTEMPTED;
Andy Adamson0382b742011-03-03 15:13:45 +0000622}
623
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000624/*
Andy Adamson16b374c2010-10-20 00:18:04 -0400625 * filelayout_check_layout()
626 *
627 * Make sure layout segment parameters are sane WRT the device.
628 * At this point no generic layer initialization of the lseg has occurred,
629 * and nothing has been added to the layout_hdr cache.
630 *
631 */
632static int
633filelayout_check_layout(struct pnfs_layout_hdr *lo,
634 struct nfs4_filelayout_segment *fl,
635 struct nfs4_layoutget_res *lgr,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400636 struct nfs4_deviceid *id,
637 gfp_t gfp_flags)
Andy Adamson16b374c2010-10-20 00:18:04 -0400638{
Benny Halevya1eaecb2011-05-19 22:14:47 -0400639 struct nfs4_deviceid_node *d;
Andy Adamson16b374c2010-10-20 00:18:04 -0400640 struct nfs4_file_layout_dsaddr *dsaddr;
641 int status = -EINVAL;
Andy Adamson16b374c2010-10-20 00:18:04 -0400642
643 dprintk("--> %s\n", __func__);
644
Andy Adamson7c24d942011-06-13 18:22:38 -0400645 /* FIXME: remove this check when layout segment support is added */
646 if (lgr->range.offset != 0 ||
647 lgr->range.length != NFS4_MAX_UINT64) {
648 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
649 __func__);
650 goto out;
651 }
652
Andy Adamson16b374c2010-10-20 00:18:04 -0400653 if (fl->pattern_offset > lgr->range.offset) {
Jim Rees3b6445a2011-02-22 19:31:57 -0500654 dprintk("%s pattern_offset %lld too large\n",
Andy Adamson16b374c2010-10-20 00:18:04 -0400655 __func__, fl->pattern_offset);
656 goto out;
657 }
658
Weston Andros Adamsonc6194272014-05-15 11:56:56 -0400659 if (!fl->stripe_unit) {
Benny Halevy75247af2011-02-22 15:56:01 -0800660 dprintk("%s Invalid stripe unit (%u)\n",
Andy Adamson16b374c2010-10-20 00:18:04 -0400661 __func__, fl->stripe_unit);
662 goto out;
663 }
664
665 /* find and reference the deviceid */
Benny Halevy35c8bb52011-05-24 18:04:02 +0300666 d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
667 NFS_SERVER(lo->plh_inode)->nfs_client, id);
Benny Halevya1eaecb2011-05-19 22:14:47 -0400668 if (d == NULL) {
Trond Myklebustcd5875f2013-05-20 11:42:54 -0400669 dsaddr = filelayout_get_device_info(lo->plh_inode, id,
670 lo->plh_lc_cred, gfp_flags);
Andy Adamson16b374c2010-10-20 00:18:04 -0400671 if (dsaddr == NULL)
672 goto out;
Benny Halevya1eaecb2011-05-19 22:14:47 -0400673 } else
674 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
Trond Myklebust1dfed272012-09-18 19:51:12 -0400675 /* Found deviceid is unavailable */
676 if (filelayout_test_devid_unavailable(&dsaddr->id_node))
Andy Adamsonc47abcf2011-06-15 17:52:40 -0400677 goto out_put;
678
Andy Adamson16b374c2010-10-20 00:18:04 -0400679 fl->dsaddr = dsaddr;
680
Chuck Levere4149662011-10-25 12:18:03 -0400681 if (fl->first_stripe_index >= dsaddr->stripe_count) {
682 dprintk("%s Bad first_stripe_index %u\n",
Andy Adamson16b374c2010-10-20 00:18:04 -0400683 __func__, fl->first_stripe_index);
684 goto out_put;
685 }
686
687 if ((fl->stripe_type == STRIPE_SPARSE &&
688 fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
689 (fl->stripe_type == STRIPE_DENSE &&
690 fl->num_fh != dsaddr->stripe_count)) {
691 dprintk("%s num_fh %u not valid for given packing\n",
692 __func__, fl->num_fh);
693 goto out_put;
694 }
695
Andy Adamson16b374c2010-10-20 00:18:04 -0400696 status = 0;
697out:
698 dprintk("--> %s returns %d\n", __func__, status);
699 return status;
700out_put:
Christoph Hellwigea8eecd2011-03-01 01:34:21 +0000701 nfs4_fl_put_deviceid(dsaddr);
Andy Adamson16b374c2010-10-20 00:18:04 -0400702 goto out;
703}
704
705static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
706{
707 int i;
708
709 for (i = 0; i < fl->num_fh; i++) {
710 if (!fl->fh_array[i])
711 break;
712 kfree(fl->fh_array[i]);
713 }
714 kfree(fl->fh_array);
715 fl->fh_array = NULL;
716}
717
718static void
719_filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
720{
721 filelayout_free_fh_array(fl);
722 kfree(fl);
723}
724
725static int
726filelayout_decode_layout(struct pnfs_layout_hdr *flo,
727 struct nfs4_filelayout_segment *fl,
728 struct nfs4_layoutget_res *lgr,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400729 struct nfs4_deviceid *id,
730 gfp_t gfp_flags)
Andy Adamson16b374c2010-10-20 00:18:04 -0400731{
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400732 struct xdr_stream stream;
Benny Halevyf7da7a12011-05-19 14:16:47 -0400733 struct xdr_buf buf;
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400734 struct page *scratch;
735 __be32 *p;
Andy Adamson16b374c2010-10-20 00:18:04 -0400736 uint32_t nfl_util;
737 int i;
738
739 dprintk("%s: set_layout_map Begin\n", __func__);
740
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400741 scratch = alloc_page(gfp_flags);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400742 if (!scratch)
743 return -ENOMEM;
744
Benny Halevyf7da7a12011-05-19 14:16:47 -0400745 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400746 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
747
748 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
749 * num_fh (4) */
750 p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
751 if (unlikely(!p))
752 goto out_err;
753
Andy Adamson16b374c2010-10-20 00:18:04 -0400754 memcpy(id, p, sizeof(*id));
755 p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
Benny Halevya1eaecb2011-05-19 22:14:47 -0400756 nfs4_print_deviceid(id);
Andy Adamson16b374c2010-10-20 00:18:04 -0400757
758 nfl_util = be32_to_cpup(p++);
759 if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
760 fl->commit_through_mds = 1;
761 if (nfl_util & NFL4_UFLG_DENSE)
762 fl->stripe_type = STRIPE_DENSE;
763 else
764 fl->stripe_type = STRIPE_SPARSE;
765 fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
766
767 fl->first_stripe_index = be32_to_cpup(p++);
768 p = xdr_decode_hyper(p, &fl->pattern_offset);
769 fl->num_fh = be32_to_cpup(p++);
770
771 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
772 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
773 fl->pattern_offset);
774
Andy Adamsoncec765c2011-06-13 18:36:17 -0400775 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
776 * Futher checking is done in filelayout_check_layout */
Chuck Levere4149662011-10-25 12:18:03 -0400777 if (fl->num_fh >
Andy Adamsoncec765c2011-06-13 18:36:17 -0400778 max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400779 goto out_err;
780
Andy Adamsoncec765c2011-06-13 18:36:17 -0400781 if (fl->num_fh > 0) {
Trond Myklebust1813bad2012-10-11 14:36:52 -0400782 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
Andy Adamsoncec765c2011-06-13 18:36:17 -0400783 gfp_flags);
784 if (!fl->fh_array)
785 goto out_err;
786 }
Andy Adamson16b374c2010-10-20 00:18:04 -0400787
788 for (i = 0; i < fl->num_fh; i++) {
789 /* Do we want to use a mempool here? */
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400790 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400791 if (!fl->fh_array[i])
792 goto out_err_free;
793
794 p = xdr_inline_decode(&stream, 4);
795 if (unlikely(!p))
796 goto out_err_free;
Andy Adamson16b374c2010-10-20 00:18:04 -0400797 fl->fh_array[i]->size = be32_to_cpup(p++);
798 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
Weston Andros Adamsonf9fd2d92012-01-26 13:32:22 -0500799 printk(KERN_ERR "NFS: Too big fh %d received %d\n",
Andy Adamson16b374c2010-10-20 00:18:04 -0400800 i, fl->fh_array[i]->size);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400801 goto out_err_free;
Andy Adamson16b374c2010-10-20 00:18:04 -0400802 }
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400803
804 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
805 if (unlikely(!p))
806 goto out_err_free;
Andy Adamson16b374c2010-10-20 00:18:04 -0400807 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
Andy Adamson16b374c2010-10-20 00:18:04 -0400808 dprintk("DEBUG: %s: fh len %d\n", __func__,
809 fl->fh_array[i]->size);
810 }
811
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400812 __free_page(scratch);
Andy Adamson16b374c2010-10-20 00:18:04 -0400813 return 0;
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400814
815out_err_free:
816 filelayout_free_fh_array(fl);
817out_err:
818 __free_page(scratch);
819 return -EIO;
Andy Adamson16b374c2010-10-20 00:18:04 -0400820}
821
Fred Isamanc8795132011-03-23 13:27:49 +0000822static void
823filelayout_free_lseg(struct pnfs_layout_segment *lseg)
824{
825 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
826
827 dprintk("--> %s\n", __func__);
828 nfs4_fl_put_deviceid(fl->dsaddr);
Fred Isaman799ba8d2012-04-20 14:47:38 -0400829 /* This assumes a single RW lseg */
830 if (lseg->pls_range.iomode == IOMODE_RW) {
831 struct nfs4_filelayout *flo;
832
833 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
834 flo->commit_info.nbuckets = 0;
835 kfree(flo->commit_info.buckets);
836 flo->commit_info.buckets = NULL;
837 }
Fred Isamanc8795132011-03-23 13:27:49 +0000838 _filelayout_free_lseg(fl);
839}
840
Fred Isaman799ba8d2012-04-20 14:47:38 -0400841static int
842filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
Fred Isamanea2cf222012-04-20 14:47:53 -0400843 struct nfs_commit_info *cinfo,
Fred Isaman799ba8d2012-04-20 14:47:38 -0400844 gfp_t gfp_flags)
845{
846 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
Fred Isamanea2cf222012-04-20 14:47:53 -0400847 struct pnfs_commit_bucket *buckets;
Weston Andros Adamsondd7663e2014-05-15 11:56:49 -0400848 int size, i;
Fred Isaman799ba8d2012-04-20 14:47:38 -0400849
850 if (fl->commit_through_mds)
851 return 0;
Weston Andros Adamsondd7663e2014-05-15 11:56:49 -0400852
853 size = (fl->stripe_type == STRIPE_SPARSE) ?
854 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
855
856 if (cinfo->ds->nbuckets >= size) {
Fred Isaman799ba8d2012-04-20 14:47:38 -0400857 /* This assumes there is only one IOMODE_RW lseg. What
858 * we really want to do is have a layout_hdr level
859 * dictionary of <multipath_list4, fh> keys, each
860 * associated with a struct list_head, populated by calls
861 * to filelayout_write_pagelist().
862 * */
863 return 0;
864 }
865
Fred Isamanea2cf222012-04-20 14:47:53 -0400866 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
Fred Isaman799ba8d2012-04-20 14:47:38 -0400867 gfp_flags);
868 if (!buckets)
869 return -ENOMEM;
Weston Andros Adamsondd7663e2014-05-15 11:56:49 -0400870 for (i = 0; i < size; i++) {
871 INIT_LIST_HEAD(&buckets[i].written);
872 INIT_LIST_HEAD(&buckets[i].committing);
Weston Andros Adamson5002c582014-05-15 11:56:54 -0400873 /* mark direct verifier as unset */
874 buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW;
Fred Isaman799ba8d2012-04-20 14:47:38 -0400875 }
Weston Andros Adamsondd7663e2014-05-15 11:56:49 -0400876
877 spin_lock(cinfo->lock);
878 if (cinfo->ds->nbuckets >= size)
879 goto out;
880 for (i = 0; i < cinfo->ds->nbuckets; i++) {
881 list_splice(&cinfo->ds->buckets[i].written,
882 &buckets[i].written);
883 list_splice(&cinfo->ds->buckets[i].committing,
884 &buckets[i].committing);
Weston Andros Adamson5002c582014-05-15 11:56:54 -0400885 buckets[i].direct_verf.committed =
886 cinfo->ds->buckets[i].direct_verf.committed;
Weston Andros Adamsondd7663e2014-05-15 11:56:49 -0400887 buckets[i].wlseg = cinfo->ds->buckets[i].wlseg;
888 buckets[i].clseg = cinfo->ds->buckets[i].clseg;
889 }
890 swap(cinfo->ds->buckets, buckets);
891 cinfo->ds->nbuckets = size;
892out:
893 spin_unlock(cinfo->lock);
894 kfree(buckets);
895 return 0;
Fred Isaman799ba8d2012-04-20 14:47:38 -0400896}
897
Andy Adamson16b374c2010-10-20 00:18:04 -0400898static struct pnfs_layout_segment *
899filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400900 struct nfs4_layoutget_res *lgr,
901 gfp_t gfp_flags)
Andy Adamson16b374c2010-10-20 00:18:04 -0400902{
903 struct nfs4_filelayout_segment *fl;
904 int rc;
905 struct nfs4_deviceid id;
906
907 dprintk("--> %s\n", __func__);
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400908 fl = kzalloc(sizeof(*fl), gfp_flags);
Andy Adamson16b374c2010-10-20 00:18:04 -0400909 if (!fl)
910 return NULL;
911
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400912 rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
913 if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
Andy Adamson16b374c2010-10-20 00:18:04 -0400914 _filelayout_free_lseg(fl);
915 return NULL;
916 }
917 return &fl->generic_hdr;
918}
919
Fred Isaman94ad1c82011-03-01 01:34:14 +0000920/*
921 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
922 *
Weston Andros Adamsonb4fdac12014-05-15 11:56:43 -0400923 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
924 * of bytes (maximum @req->wb_bytes) that can be coalesced.
Fred Isaman94ad1c82011-03-01 01:34:14 +0000925 */
Weston Andros Adamsonb4fdac12014-05-15 11:56:43 -0400926static size_t
Fred Isaman94ad1c82011-03-01 01:34:14 +0000927filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
928 struct nfs_page *req)
929{
Weston Andros Adamson0f9c4292014-05-15 11:56:51 -0400930 unsigned int size;
Fred Isaman94ad1c82011-03-01 01:34:14 +0000931 u64 p_stripe, r_stripe;
Weston Andros Adamsonc6194272014-05-15 11:56:56 -0400932 u32 stripe_offset;
933 u64 segment_offset = pgio->pg_lseg->pls_range.offset;
934 u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
Fred Isaman94ad1c82011-03-01 01:34:14 +0000935
Weston Andros Adamson0f9c4292014-05-15 11:56:51 -0400936 /* calls nfs_generic_pg_test */
937 size = pnfs_generic_pg_test(pgio, prev, req);
938 if (!size)
Weston Andros Adamsonb4fdac12014-05-15 11:56:43 -0400939 return 0;
Benny Halevy89a58e32011-05-25 20:54:40 +0300940
Weston Andros Adamsonc6194272014-05-15 11:56:56 -0400941 /* see if req and prev are in the same stripe */
Weston Andros Adamson0f9c4292014-05-15 11:56:51 -0400942 if (prev) {
Weston Andros Adamsonc6194272014-05-15 11:56:56 -0400943 p_stripe = (u64)req_offset(prev) - segment_offset;
944 r_stripe = (u64)req_offset(req) - segment_offset;
Weston Andros Adamson0f9c4292014-05-15 11:56:51 -0400945 do_div(p_stripe, stripe_unit);
946 do_div(r_stripe, stripe_unit);
Fred Isaman94ad1c82011-03-01 01:34:14 +0000947
Weston Andros Adamson0f9c4292014-05-15 11:56:51 -0400948 if (p_stripe != r_stripe)
949 return 0;
950 }
Weston Andros Adamsonc6194272014-05-15 11:56:56 -0400951
952 /* calculate remaining bytes in the current stripe */
953 div_u64_rem((u64)req_offset(req) - segment_offset,
954 stripe_unit,
955 &stripe_offset);
956 WARN_ON_ONCE(stripe_offset > stripe_unit);
957 if (stripe_offset >= stripe_unit)
958 return 0;
959 return min(stripe_unit - (unsigned int)stripe_offset, size);
Fred Isaman94ad1c82011-03-01 01:34:14 +0000960}
961
Trond Myklebust17280172012-03-11 13:11:00 -0400962static void
Andy Adamson7c24d942011-06-13 18:22:38 -0400963filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
964 struct nfs_page *req)
965{
Weston Andros Adamsonc6194272014-05-15 11:56:56 -0400966 if (!pgio->pg_lseg)
967 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
Andy Adamson7c24d942011-06-13 18:22:38 -0400968 req->wb_context,
969 0,
970 NFS4_MAX_UINT64,
971 IOMODE_READ,
972 GFP_KERNEL);
973 /* If no lseg, fall back to read through mds */
974 if (pgio->pg_lseg == NULL)
Trond Myklebust1f945352011-07-13 15:59:57 -0400975 nfs_pageio_reset_read_mds(pgio);
Andy Adamson7c24d942011-06-13 18:22:38 -0400976}
977
Trond Myklebust17280172012-03-11 13:11:00 -0400978static void
Andy Adamson7c24d942011-06-13 18:22:38 -0400979filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
980 struct nfs_page *req)
981{
Fred Isamanea2cf222012-04-20 14:47:53 -0400982 struct nfs_commit_info cinfo;
Fred Isaman799ba8d2012-04-20 14:47:38 -0400983 int status;
984
Weston Andros Adamsonc6194272014-05-15 11:56:56 -0400985 if (!pgio->pg_lseg)
986 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
Andy Adamson7c24d942011-06-13 18:22:38 -0400987 req->wb_context,
988 0,
989 NFS4_MAX_UINT64,
990 IOMODE_RW,
991 GFP_NOFS);
992 /* If no lseg, fall back to write through mds */
993 if (pgio->pg_lseg == NULL)
Fred Isaman799ba8d2012-04-20 14:47:38 -0400994 goto out_mds;
Fred Isamanea2cf222012-04-20 14:47:53 -0400995 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
996 status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
Fred Isaman799ba8d2012-04-20 14:47:38 -0400997 if (status < 0) {
Trond Myklebust9369a432012-09-18 20:57:08 -0400998 pnfs_put_lseg(pgio->pg_lseg);
Fred Isaman799ba8d2012-04-20 14:47:38 -0400999 pgio->pg_lseg = NULL;
1000 goto out_mds;
1001 }
1002 return;
1003out_mds:
1004 nfs_pageio_reset_write_mds(pgio);
Andy Adamson7c24d942011-06-13 18:22:38 -04001005}
1006
Trond Myklebust1751c362011-06-10 13:30:23 -04001007static const struct nfs_pageio_ops filelayout_pg_read_ops = {
Andy Adamson7c24d942011-06-13 18:22:38 -04001008 .pg_init = filelayout_pg_init_read,
Trond Myklebust1751c362011-06-10 13:30:23 -04001009 .pg_test = filelayout_pg_test,
Trond Myklebust493292d2011-07-13 15:58:28 -04001010 .pg_doio = pnfs_generic_pg_readpages,
Trond Myklebust1751c362011-06-10 13:30:23 -04001011};
1012
1013static const struct nfs_pageio_ops filelayout_pg_write_ops = {
Andy Adamson7c24d942011-06-13 18:22:38 -04001014 .pg_init = filelayout_pg_init_write,
Trond Myklebust1751c362011-06-10 13:30:23 -04001015 .pg_test = filelayout_pg_test,
Trond Myklebustdce81292011-07-13 15:59:19 -04001016 .pg_doio = pnfs_generic_pg_writepages,
Trond Myklebust1751c362011-06-10 13:30:23 -04001017};
1018
Fred Isamane0c2b382011-03-23 13:27:53 +00001019static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
1020{
1021 if (fl->stripe_type == STRIPE_SPARSE)
1022 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
1023 else
1024 return j;
1025}
1026
Fred Isamand6d6dc72012-03-08 17:29:35 -05001027/* The generic layer is about to remove the req from the commit list.
1028 * If this will make the bucket empty, it will need to put the lseg reference.
Fred Isamand6d6dc72012-03-08 17:29:35 -05001029 */
Trond Myklebust8dd37752012-03-15 17:16:40 -04001030static void
Fred Isamanea2cf222012-04-20 14:47:53 -04001031filelayout_clear_request_commit(struct nfs_page *req,
1032 struct nfs_commit_info *cinfo)
Fred Isamane0c2b382011-03-23 13:27:53 +00001033{
Trond Myklebust8dd37752012-03-15 17:16:40 -04001034 struct pnfs_layout_segment *freeme = NULL;
Trond Myklebust8dd37752012-03-15 17:16:40 -04001035
Fred Isamanea2cf222012-04-20 14:47:53 -04001036 spin_lock(cinfo->lock);
Trond Myklebust8dd37752012-03-15 17:16:40 -04001037 if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
1038 goto out;
Fred Isamanea2cf222012-04-20 14:47:53 -04001039 cinfo->ds->nwritten--;
Fred Isamand6d6dc72012-03-08 17:29:35 -05001040 if (list_is_singular(&req->wb_list)) {
Fred Isamanea2cf222012-04-20 14:47:53 -04001041 struct pnfs_commit_bucket *bucket;
Fred Isamand6d6dc72012-03-08 17:29:35 -05001042
Fred Isaman799ba8d2012-04-20 14:47:38 -04001043 bucket = list_first_entry(&req->wb_list,
Fred Isamanea2cf222012-04-20 14:47:53 -04001044 struct pnfs_commit_bucket,
Fred Isaman799ba8d2012-04-20 14:47:38 -04001045 written);
1046 freeme = bucket->wlseg;
1047 bucket->wlseg = NULL;
Fred Isamand6d6dc72012-03-08 17:29:35 -05001048 }
Trond Myklebust8dd37752012-03-15 17:16:40 -04001049out:
Fred Isamanea2cf222012-04-20 14:47:53 -04001050 nfs_request_remove_commit_list(req, cinfo);
1051 spin_unlock(cinfo->lock);
Trond Myklebust9369a432012-09-18 20:57:08 -04001052 pnfs_put_lseg(freeme);
Fred Isamand6d6dc72012-03-08 17:29:35 -05001053}
1054
1055static struct list_head *
1056filelayout_choose_commit_list(struct nfs_page *req,
Fred Isamanea2cf222012-04-20 14:47:53 -04001057 struct pnfs_layout_segment *lseg,
1058 struct nfs_commit_info *cinfo)
Fred Isamand6d6dc72012-03-08 17:29:35 -05001059{
Fred Isamane0c2b382011-03-23 13:27:53 +00001060 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1061 u32 i, j;
1062 struct list_head *list;
Fred Isamanea2cf222012-04-20 14:47:53 -04001063 struct pnfs_commit_bucket *buckets;
Fred Isamane0c2b382011-03-23 13:27:53 +00001064
Fred Isamand6d6dc72012-03-08 17:29:35 -05001065 if (fl->commit_through_mds)
Fred Isamanea2cf222012-04-20 14:47:53 -04001066 return &cinfo->mds->list;
Fred Isamand6d6dc72012-03-08 17:29:35 -05001067
Fred Isamane0c2b382011-03-23 13:27:53 +00001068 /* Note that we are calling nfs4_fl_calc_j_index on each page
1069 * that ends up being committed to a data server. An attractive
1070 * alternative is to add a field to nfs_write_data and nfs_page
1071 * to store the value calculated in filelayout_write_pagelist
1072 * and just use that here.
1073 */
Fred Isamanb5542842012-04-20 14:47:43 -04001074 j = nfs4_fl_calc_j_index(lseg, req_offset(req));
Fred Isamane0c2b382011-03-23 13:27:53 +00001075 i = select_bucket_index(fl, j);
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001076 spin_lock(cinfo->lock);
Fred Isamanea2cf222012-04-20 14:47:53 -04001077 buckets = cinfo->ds->buckets;
Fred Isaman799ba8d2012-04-20 14:47:38 -04001078 list = &buckets[i].written;
Fred Isamane0c2b382011-03-23 13:27:53 +00001079 if (list_empty(list)) {
Fred Isamand6d6dc72012-03-08 17:29:35 -05001080 /* Non-empty buckets hold a reference on the lseg. That ref
1081 * is normally transferred to the COMMIT call and released
1082 * there. It could also be released if the last req is pulled
1083 * off due to a rewrite, in which case it will be done in
Fred Isaman799ba8d2012-04-20 14:47:38 -04001084 * filelayout_clear_request_commit
Fred Isamand6d6dc72012-03-08 17:29:35 -05001085 */
Trond Myklebust9369a432012-09-18 20:57:08 -04001086 buckets[i].wlseg = pnfs_get_lseg(lseg);
Fred Isamane0c2b382011-03-23 13:27:53 +00001087 }
Trond Myklebust8dd37752012-03-15 17:16:40 -04001088 set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
Fred Isamanea2cf222012-04-20 14:47:53 -04001089 cinfo->ds->nwritten++;
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001090 spin_unlock(cinfo->lock);
Fred Isamane0c2b382011-03-23 13:27:53 +00001091 return list;
1092}
1093
Trond Myklebust8dd37752012-03-15 17:16:40 -04001094static void
1095filelayout_mark_request_commit(struct nfs_page *req,
Fred Isamanea2cf222012-04-20 14:47:53 -04001096 struct pnfs_layout_segment *lseg,
1097 struct nfs_commit_info *cinfo)
Trond Myklebust8dd37752012-03-15 17:16:40 -04001098{
1099 struct list_head *list;
1100
Fred Isamanea2cf222012-04-20 14:47:53 -04001101 list = filelayout_choose_commit_list(req, lseg, cinfo);
1102 nfs_request_add_commit_list(req, list, cinfo);
Trond Myklebust8dd37752012-03-15 17:16:40 -04001103}
1104
Fred Isamane0c2b382011-03-23 13:27:53 +00001105static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1106{
1107 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1108
1109 if (flseg->stripe_type == STRIPE_SPARSE)
1110 return i;
1111 else
1112 return nfs4_fl_calc_ds_index(lseg, i);
1113}
1114
1115static struct nfs_fh *
1116select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1117{
1118 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1119
1120 if (flseg->stripe_type == STRIPE_SPARSE) {
1121 if (flseg->num_fh == 1)
1122 i = 0;
1123 else if (flseg->num_fh == 0)
1124 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1125 return NULL;
1126 }
1127 return flseg->fh_array[i];
1128}
1129
Fred Isaman0b7c0152012-04-20 14:47:39 -04001130static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
Fred Isamane0c2b382011-03-23 13:27:53 +00001131{
1132 struct pnfs_layout_segment *lseg = data->lseg;
1133 struct nfs4_pnfs_ds *ds;
Andy Adamson0e201622013-09-06 14:14:00 -04001134 struct rpc_clnt *ds_clnt;
Fred Isamane0c2b382011-03-23 13:27:53 +00001135 u32 idx;
1136 struct nfs_fh *fh;
1137
1138 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1139 ds = nfs4_fl_prepare_ds(lseg, idx);
Andy Adamson0e201622013-09-06 14:14:00 -04001140 if (!ds)
1141 goto out_err;
1142
1143 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1144 if (IS_ERR(ds_clnt))
1145 goto out_err;
1146
Andy Adamson3a7936c2012-04-27 17:53:51 -04001147 dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1148 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
Fred Isaman0b7c0152012-04-20 14:47:39 -04001149 data->commit_done_cb = filelayout_commit_done_cb;
Andy Adamson3a7936c2012-04-27 17:53:51 -04001150 atomic_inc(&ds->ds_clp->cl_count);
Fred Isamane0c2b382011-03-23 13:27:53 +00001151 data->ds_clp = ds->ds_clp;
1152 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1153 if (fh)
1154 data->args.fh = fh;
Andy Adamson0e201622013-09-06 14:14:00 -04001155 return nfs_initiate_commit(ds_clnt, data,
Andy Adamson9f0ec1762012-04-27 17:53:44 -04001156 &filelayout_commit_call_ops, how,
1157 RPC_TASK_SOFTCONN);
Andy Adamson0e201622013-09-06 14:14:00 -04001158out_err:
1159 prepare_to_resend_writes(data);
1160 filelayout_commit_release(data);
1161 return -EAGAIN;
Fred Isamane0c2b382011-03-23 13:27:53 +00001162}
1163
Trond Myklebust8dd37752012-03-15 17:16:40 -04001164static int
Fred Isaman1763da12012-04-20 14:47:57 -04001165transfer_commit_list(struct list_head *src, struct list_head *dst,
1166 struct nfs_commit_info *cinfo, int max)
Trond Myklebust8dd37752012-03-15 17:16:40 -04001167{
Trond Myklebust8dd37752012-03-15 17:16:40 -04001168 struct nfs_page *req, *tmp;
1169 int ret = 0;
1170
1171 list_for_each_entry_safe(req, tmp, src, wb_list) {
1172 if (!nfs_lock_request(req))
1173 continue;
Trond Myklebust53b8ee32012-05-22 16:36:27 -04001174 kref_get(&req->wb_kref);
Fred Isamanea2cf222012-04-20 14:47:53 -04001175 if (cond_resched_lock(cinfo->lock))
Trond Myklebust3b3be882012-03-17 11:59:30 -04001176 list_safe_reset_next(req, tmp, wb_list);
Fred Isamanea2cf222012-04-20 14:47:53 -04001177 nfs_request_remove_commit_list(req, cinfo);
Trond Myklebust8dd37752012-03-15 17:16:40 -04001178 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1179 nfs_list_add_request(req, dst);
1180 ret++;
Fred Isaman1763da12012-04-20 14:47:57 -04001181 if ((ret == max) && !cinfo->dreq)
Trond Myklebust8dd37752012-03-15 17:16:40 -04001182 break;
1183 }
Fred Isaman1763da12012-04-20 14:47:57 -04001184 return ret;
1185}
1186
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001187/* Note called with cinfo->lock held. */
Fred Isaman1763da12012-04-20 14:47:57 -04001188static int
1189filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1190 struct nfs_commit_info *cinfo,
1191 int max)
1192{
1193 struct list_head *src = &bucket->written;
1194 struct list_head *dst = &bucket->committing;
1195 int ret;
1196
1197 ret = transfer_commit_list(src, dst, cinfo, max);
Fred Isaman799ba8d2012-04-20 14:47:38 -04001198 if (ret) {
Fred Isamanea2cf222012-04-20 14:47:53 -04001199 cinfo->ds->nwritten -= ret;
1200 cinfo->ds->ncommitting += ret;
Fred Isaman799ba8d2012-04-20 14:47:38 -04001201 bucket->clseg = bucket->wlseg;
1202 if (list_empty(src))
1203 bucket->wlseg = NULL;
1204 else
Trond Myklebust9369a432012-09-18 20:57:08 -04001205 pnfs_get_lseg(bucket->clseg);
Fred Isaman799ba8d2012-04-20 14:47:38 -04001206 }
Trond Myklebust8dd37752012-03-15 17:16:40 -04001207 return ret;
1208}
1209
Fred Isamand6d6dc72012-03-08 17:29:35 -05001210/* Move reqs from written to committing lists, returning count of number moved.
Fred Isamanea2cf222012-04-20 14:47:53 -04001211 * Note called with cinfo->lock held.
Fred Isamand6d6dc72012-03-08 17:29:35 -05001212 */
Fred Isamanea2cf222012-04-20 14:47:53 -04001213static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1214 int max)
Fred Isamand6d6dc72012-03-08 17:29:35 -05001215{
Fred Isamand6d6dc72012-03-08 17:29:35 -05001216 int i, rv = 0, cnt;
1217
Fred Isamanea2cf222012-04-20 14:47:53 -04001218 for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1219 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1220 cinfo, max);
Fred Isamand6d6dc72012-03-08 17:29:35 -05001221 max -= cnt;
1222 rv += cnt;
1223 }
Fred Isamand6d6dc72012-03-08 17:29:35 -05001224 return rv;
1225}
1226
Fred Isaman1763da12012-04-20 14:47:57 -04001227/* Pull everything off the committing lists and dump into @dst */
1228static void filelayout_recover_commit_reqs(struct list_head *dst,
1229 struct nfs_commit_info *cinfo)
1230{
1231 struct pnfs_commit_bucket *b;
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001232 struct pnfs_layout_segment *freeme;
Fred Isaman1763da12012-04-20 14:47:57 -04001233 int i;
1234
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001235restart:
Weston Andros Adamson471252c2014-01-21 15:21:33 -05001236 spin_lock(cinfo->lock);
Fred Isaman1763da12012-04-20 14:47:57 -04001237 for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1238 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001239 freeme = b->wlseg;
Fred Isaman1763da12012-04-20 14:47:57 -04001240 b->wlseg = NULL;
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001241 spin_unlock(cinfo->lock);
1242 pnfs_put_lseg(freeme);
1243 goto restart;
Fred Isaman1763da12012-04-20 14:47:57 -04001244 }
1245 }
1246 cinfo->ds->nwritten = 0;
Weston Andros Adamson471252c2014-01-21 15:21:33 -05001247 spin_unlock(cinfo->lock);
Fred Isaman1763da12012-04-20 14:47:57 -04001248}
1249
Trond Myklebust9390f422012-03-16 13:52:45 -04001250static unsigned int
Fred Isamanea2cf222012-04-20 14:47:53 -04001251alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
Fred Isamane0c2b382011-03-23 13:27:53 +00001252{
Fred Isamanea2cf222012-04-20 14:47:53 -04001253 struct pnfs_ds_commit_info *fl_cinfo;
1254 struct pnfs_commit_bucket *bucket;
Fred Isaman0b7c0152012-04-20 14:47:39 -04001255 struct nfs_commit_data *data;
Fred Isamane0c2b382011-03-23 13:27:53 +00001256 int i, j;
Trond Myklebust9390f422012-03-16 13:52:45 -04001257 unsigned int nreq = 0;
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001258 struct pnfs_layout_segment *freeme;
Fred Isamane0c2b382011-03-23 13:27:53 +00001259
Fred Isamanea2cf222012-04-20 14:47:53 -04001260 fl_cinfo = cinfo->ds;
Fred Isaman799ba8d2012-04-20 14:47:38 -04001261 bucket = fl_cinfo->buckets;
1262 for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1263 if (list_empty(&bucket->committing))
Fred Isamane0c2b382011-03-23 13:27:53 +00001264 continue;
1265 data = nfs_commitdata_alloc();
1266 if (!data)
Trond Myklebust9390f422012-03-16 13:52:45 -04001267 break;
Fred Isamane0c2b382011-03-23 13:27:53 +00001268 data->ds_commit_index = i;
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001269 spin_lock(cinfo->lock);
Fred Isaman799ba8d2012-04-20 14:47:38 -04001270 data->lseg = bucket->clseg;
1271 bucket->clseg = NULL;
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001272 spin_unlock(cinfo->lock);
Fred Isamane0c2b382011-03-23 13:27:53 +00001273 list_add(&data->pages, list);
Trond Myklebust9390f422012-03-16 13:52:45 -04001274 nreq++;
Fred Isamane0c2b382011-03-23 13:27:53 +00001275 }
Fred Isamane0c2b382011-03-23 13:27:53 +00001276
Trond Myklebust9390f422012-03-16 13:52:45 -04001277 /* Clean up on error */
Fred Isaman799ba8d2012-04-20 14:47:38 -04001278 for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1279 if (list_empty(&bucket->committing))
Fred Isamane0c2b382011-03-23 13:27:53 +00001280 continue;
Fred Isamanea2cf222012-04-20 14:47:53 -04001281 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001282 spin_lock(cinfo->lock);
1283 freeme = bucket->clseg;
Fred Isaman799ba8d2012-04-20 14:47:38 -04001284 bucket->clseg = NULL;
Weston Andros Adamsond201c4d2014-05-15 11:56:40 -04001285 spin_unlock(cinfo->lock);
1286 pnfs_put_lseg(freeme);
Fred Isamane0c2b382011-03-23 13:27:53 +00001287 }
Fred Isamane0c2b382011-03-23 13:27:53 +00001288 /* Caller will clean up entries put on list */
Trond Myklebust9390f422012-03-16 13:52:45 -04001289 return nreq;
Fred Isamane0c2b382011-03-23 13:27:53 +00001290}
1291
1292/* This follows nfs_commit_list pretty closely */
1293static int
1294filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
Fred Isamanea2cf222012-04-20 14:47:53 -04001295 int how, struct nfs_commit_info *cinfo)
Fred Isamane0c2b382011-03-23 13:27:53 +00001296{
Fred Isaman0b7c0152012-04-20 14:47:39 -04001297 struct nfs_commit_data *data, *tmp;
Fred Isamane0c2b382011-03-23 13:27:53 +00001298 LIST_HEAD(list);
Trond Myklebust9390f422012-03-16 13:52:45 -04001299 unsigned int nreq = 0;
Fred Isamane0c2b382011-03-23 13:27:53 +00001300
1301 if (!list_empty(mds_pages)) {
1302 data = nfs_commitdata_alloc();
Trond Myklebust9390f422012-03-16 13:52:45 -04001303 if (data != NULL) {
1304 data->lseg = NULL;
1305 list_add(&data->pages, &list);
1306 nreq++;
1307 } else
Fred Isamanea2cf222012-04-20 14:47:53 -04001308 nfs_retry_commit(mds_pages, NULL, cinfo);
Fred Isamane0c2b382011-03-23 13:27:53 +00001309 }
1310
Fred Isamanea2cf222012-04-20 14:47:53 -04001311 nreq += alloc_ds_commits(cinfo, &list);
Trond Myklebust9390f422012-03-16 13:52:45 -04001312
1313 if (nreq == 0) {
Fred Isamanf453a542012-04-20 14:47:54 -04001314 cinfo->completion_ops->error_cleanup(NFS_I(inode));
Trond Myklebust9390f422012-03-16 13:52:45 -04001315 goto out;
1316 }
1317
Fred Isamanea2cf222012-04-20 14:47:53 -04001318 atomic_add(nreq, &cinfo->mds->rpcs_out);
Fred Isamane0c2b382011-03-23 13:27:53 +00001319
1320 list_for_each_entry_safe(data, tmp, &list, pages) {
1321 list_del_init(&data->pages);
Fred Isamane0c2b382011-03-23 13:27:53 +00001322 if (!data->lseg) {
Fred Isamanf453a542012-04-20 14:47:54 -04001323 nfs_init_commit(data, mds_pages, NULL, cinfo);
Fred Isaman0b7c0152012-04-20 14:47:39 -04001324 nfs_initiate_commit(NFS_CLIENT(inode), data,
Andy Adamson9f0ec1762012-04-27 17:53:44 -04001325 data->mds_ops, how, 0);
Fred Isamane0c2b382011-03-23 13:27:53 +00001326 } else {
Fred Isamanea2cf222012-04-20 14:47:53 -04001327 struct pnfs_commit_bucket *buckets;
Fred Isaman799ba8d2012-04-20 14:47:38 -04001328
Fred Isamanea2cf222012-04-20 14:47:53 -04001329 buckets = cinfo->ds->buckets;
Fred Isamanf453a542012-04-20 14:47:54 -04001330 nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
Fred Isamane0c2b382011-03-23 13:27:53 +00001331 filelayout_initiate_commit(data, how);
1332 }
1333 }
Trond Myklebust9390f422012-03-16 13:52:45 -04001334out:
Fred Isamanea2cf222012-04-20 14:47:53 -04001335 cinfo->ds->ncommitting = 0;
Trond Myklebust9390f422012-03-16 13:52:45 -04001336 return PNFS_ATTEMPTED;
Fred Isamane0c2b382011-03-23 13:27:53 +00001337}
1338
Benny Halevy1775bc32011-05-20 13:47:33 +02001339static void
1340filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1341{
1342 nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1343}
1344
Fred Isaman799ba8d2012-04-20 14:47:38 -04001345static struct pnfs_layout_hdr *
1346filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1347{
1348 struct nfs4_filelayout *flo;
1349
1350 flo = kzalloc(sizeof(*flo), gfp_flags);
Trond Myklebust6df200f2014-05-29 20:06:55 -04001351 return flo != NULL ? &flo->generic_hdr : NULL;
Fred Isaman799ba8d2012-04-20 14:47:38 -04001352}
1353
1354static void
1355filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1356{
1357 kfree(FILELAYOUT_FROM_HDR(lo));
1358}
1359
Fred Isamanea2cf222012-04-20 14:47:53 -04001360static struct pnfs_ds_commit_info *
1361filelayout_get_ds_info(struct inode *inode)
1362{
Fred Isamandf011742012-04-24 14:50:34 -04001363 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1364
1365 if (layout == NULL)
1366 return NULL;
1367 else
1368 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
Fred Isamanea2cf222012-04-20 14:47:53 -04001369}
1370
Dean Hildebrand7ab672c2010-10-20 00:18:00 -04001371static struct pnfs_layoutdriver_type filelayout_type = {
Christoph Hellwigea8eecd2011-03-01 01:34:21 +00001372 .id = LAYOUT_NFSV4_1_FILES,
1373 .name = "LAYOUT_NFSV4_1_FILES",
1374 .owner = THIS_MODULE,
Fred Isaman799ba8d2012-04-20 14:47:38 -04001375 .alloc_layout_hdr = filelayout_alloc_layout_hdr,
1376 .free_layout_hdr = filelayout_free_layout_hdr,
Christoph Hellwigea8eecd2011-03-01 01:34:21 +00001377 .alloc_lseg = filelayout_alloc_lseg,
1378 .free_lseg = filelayout_free_lseg,
Trond Myklebust1751c362011-06-10 13:30:23 -04001379 .pg_read_ops = &filelayout_pg_read_ops,
1380 .pg_write_ops = &filelayout_pg_write_ops,
Fred Isamanea2cf222012-04-20 14:47:53 -04001381 .get_ds_info = &filelayout_get_ds_info,
Trond Myklebust8dd37752012-03-15 17:16:40 -04001382 .mark_request_commit = filelayout_mark_request_commit,
1383 .clear_request_commit = filelayout_clear_request_commit,
Fred Isamand6d6dc72012-03-08 17:29:35 -05001384 .scan_commit_lists = filelayout_scan_commit_lists,
Fred Isaman1763da12012-04-20 14:47:57 -04001385 .recover_commit_reqs = filelayout_recover_commit_reqs,
Fred Isamane0c2b382011-03-23 13:27:53 +00001386 .commit_pagelist = filelayout_commit_pagelist,
Andy Adamsondc70d7b2011-03-01 01:34:19 +00001387 .read_pagelist = filelayout_read_pagelist,
Andy Adamson0382b742011-03-03 15:13:45 +00001388 .write_pagelist = filelayout_write_pagelist,
Benny Halevy1775bc32011-05-20 13:47:33 +02001389 .free_deviceid_node = filelayout_free_deveiceid_node,
Dean Hildebrand7ab672c2010-10-20 00:18:00 -04001390};
1391
1392static int __init nfs4filelayout_init(void)
1393{
1394 printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1395 __func__);
1396 return pnfs_register_layoutdriver(&filelayout_type);
1397}
1398
1399static void __exit nfs4filelayout_exit(void)
1400{
1401 printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1402 __func__);
1403 pnfs_unregister_layoutdriver(&filelayout_type);
1404}
1405
J. Bruce Fieldsf85ef692011-07-15 19:18:42 -04001406MODULE_ALIAS("nfs-layouttype4-1");
1407
Dean Hildebrand7ab672c2010-10-20 00:18:00 -04001408module_init(nfs4filelayout_init);
1409module_exit(nfs4filelayout_exit);