blob: 33bda24e8cd24cdd1485d6108f31cdbafee4cf57 [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>
Andy Adamson16b374c2010-10-20 00:18:04 -040033
34#include "internal.h"
35#include "nfs4filelayout.h"
Dean Hildebrand7ab672c2010-10-20 00:18:00 -040036
37#define NFSDBG_FACILITY NFSDBG_PNFS_LD
38
39MODULE_LICENSE("GPL");
40MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
41MODULE_DESCRIPTION("The NFSv4 file layout driver");
42
Andy Adamsoncbdabc7f2011-03-01 01:34:20 +000043#define FILELAYOUT_POLL_RETRY_MAX (15*HZ)
44
Fred Isamancfe7f412011-03-01 01:34:18 +000045static loff_t
46filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
47 loff_t offset)
48{
49 u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
50 u64 tmp;
51
52 offset -= flseg->pattern_offset;
53 tmp = offset;
54 do_div(tmp, stripe_width);
55
56 return tmp * flseg->stripe_unit + do_div(offset, flseg->stripe_unit);
57}
58
59/* This function is used by the layout driver to calculate the
60 * offset of the file on the dserver based on whether the
61 * layout type is STRIPE_DENSE or STRIPE_SPARSE
62 */
63static loff_t
64filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
65{
66 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
67
68 switch (flseg->stripe_type) {
69 case STRIPE_SPARSE:
70 return offset;
71
72 case STRIPE_DENSE:
73 return filelayout_get_dense_offset(flseg, offset);
74 }
75
76 BUG();
77}
78
Andy Adamsoncbdabc7f2011-03-01 01:34:20 +000079/* For data server errors we don't recover from */
80static void
81filelayout_set_lo_fail(struct pnfs_layout_segment *lseg)
82{
83 if (lseg->pls_range.iomode == IOMODE_RW) {
84 dprintk("%s Setting layout IOMODE_RW fail bit\n", __func__);
85 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
86 } else {
87 dprintk("%s Setting layout IOMODE_READ fail bit\n", __func__);
88 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
89 }
90}
91
92static int filelayout_async_handle_error(struct rpc_task *task,
93 struct nfs4_state *state,
94 struct nfs_client *clp,
95 int *reset)
96{
97 if (task->tk_status >= 0)
98 return 0;
99
100 *reset = 0;
101
102 switch (task->tk_status) {
103 case -NFS4ERR_BADSESSION:
104 case -NFS4ERR_BADSLOT:
105 case -NFS4ERR_BAD_HIGH_SLOT:
106 case -NFS4ERR_DEADSESSION:
107 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
108 case -NFS4ERR_SEQ_FALSE_RETRY:
109 case -NFS4ERR_SEQ_MISORDERED:
110 dprintk("%s ERROR %d, Reset session. Exchangeid "
111 "flags 0x%x\n", __func__, task->tk_status,
112 clp->cl_exchange_flags);
113 nfs4_schedule_session_recovery(clp->cl_session);
114 break;
115 case -NFS4ERR_DELAY:
116 case -NFS4ERR_GRACE:
117 case -EKEYEXPIRED:
118 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
119 break;
Andy Adamsona8a4ae32011-05-03 13:43:03 -0400120 case -NFS4ERR_RETRY_UNCACHED_REP:
121 break;
Andy Adamsoncbdabc7f2011-03-01 01:34:20 +0000122 default:
123 dprintk("%s DS error. Retry through MDS %d\n", __func__,
124 task->tk_status);
125 *reset = 1;
126 break;
127 }
128 task->tk_status = 0;
129 return -EAGAIN;
130}
131
132/* NFS_PROTO call done callback routines */
133
134static int filelayout_read_done_cb(struct rpc_task *task,
135 struct nfs_read_data *data)
136{
137 struct nfs_client *clp = data->ds_clp;
138 int reset = 0;
139
140 dprintk("%s DS read\n", __func__);
141
142 if (filelayout_async_handle_error(task, data->args.context->state,
143 data->ds_clp, &reset) == -EAGAIN) {
144 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
145 __func__, data->ds_clp, data->ds_clp->cl_session);
146 if (reset) {
147 filelayout_set_lo_fail(data->lseg);
148 nfs4_reset_read(task, data);
149 clp = NFS_SERVER(data->inode)->nfs_client;
150 }
151 nfs_restart_rpc(task, clp);
152 return -EAGAIN;
153 }
154
155 return 0;
156}
157
Andy Adamson16b374c2010-10-20 00:18:04 -0400158/*
Andy Adamson863a3c62011-03-23 13:27:54 +0000159 * We reference the rpc_cred of the first WRITE that triggers the need for
160 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
161 * rfc5661 is not clear about which credential should be used.
162 */
163static void
164filelayout_set_layoutcommit(struct nfs_write_data *wdata)
165{
166 if (FILELAYOUT_LSEG(wdata->lseg)->commit_through_mds ||
167 wdata->res.verf->committed == NFS_FILE_SYNC)
168 return;
169
170 pnfs_set_layoutcommit(wdata);
171 dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, wdata->inode->i_ino,
172 (unsigned long) wdata->lseg->pls_end_pos);
173}
174
175/*
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000176 * Call ops for the async read/write cases
177 * In the case of dense layouts, the offset needs to be reset to its
178 * original value.
179 */
180static void filelayout_read_prepare(struct rpc_task *task, void *data)
181{
182 struct nfs_read_data *rdata = (struct nfs_read_data *)data;
183
Andy Adamsoncbdabc7f2011-03-01 01:34:20 +0000184 rdata->read_done_cb = filelayout_read_done_cb;
185
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000186 if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
187 &rdata->args.seq_args, &rdata->res.seq_res,
188 0, task))
189 return;
190
191 rpc_call_start(task);
192}
193
194static void filelayout_read_call_done(struct rpc_task *task, void *data)
195{
196 struct nfs_read_data *rdata = (struct nfs_read_data *)data;
197
198 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
199
200 /* Note this may cause RPC to be resent */
201 rdata->mds_ops->rpc_call_done(task, data);
202}
203
204static void filelayout_read_release(void *data)
205{
206 struct nfs_read_data *rdata = (struct nfs_read_data *)data;
207
208 rdata->mds_ops->rpc_release(data);
209}
210
Fred Isamana69aef12011-03-03 15:13:47 +0000211static int filelayout_write_done_cb(struct rpc_task *task,
212 struct nfs_write_data *data)
213{
214 int reset = 0;
215
216 if (filelayout_async_handle_error(task, data->args.context->state,
217 data->ds_clp, &reset) == -EAGAIN) {
218 struct nfs_client *clp;
219
220 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
221 __func__, data->ds_clp, data->ds_clp->cl_session);
222 if (reset) {
223 filelayout_set_lo_fail(data->lseg);
224 nfs4_reset_write(task, data);
225 clp = NFS_SERVER(data->inode)->nfs_client;
226 } else
227 clp = data->ds_clp;
228 nfs_restart_rpc(task, clp);
229 return -EAGAIN;
230 }
231
Andy Adamson863a3c62011-03-23 13:27:54 +0000232 filelayout_set_layoutcommit(data);
Fred Isamana69aef12011-03-03 15:13:47 +0000233 return 0;
234}
235
Fred Isamane0c2b382011-03-23 13:27:53 +0000236/* Fake up some data that will cause nfs_commit_release to retry the writes. */
237static void prepare_to_resend_writes(struct nfs_write_data *data)
238{
239 struct nfs_page *first = nfs_list_entry(data->pages.next);
240
241 data->task.tk_status = 0;
242 memcpy(data->verf.verifier, first->wb_verf.verifier,
243 sizeof(first->wb_verf.verifier));
244 data->verf.verifier[0]++; /* ensure verifier mismatch */
245}
246
247static int filelayout_commit_done_cb(struct rpc_task *task,
248 struct nfs_write_data *data)
249{
250 int reset = 0;
251
252 if (filelayout_async_handle_error(task, data->args.context->state,
253 data->ds_clp, &reset) == -EAGAIN) {
254 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
255 __func__, data->ds_clp, data->ds_clp->cl_session);
256 if (reset) {
257 prepare_to_resend_writes(data);
258 filelayout_set_lo_fail(data->lseg);
259 } else
260 nfs_restart_rpc(task, data->ds_clp);
261 return -EAGAIN;
262 }
263
264 return 0;
265}
266
Fred Isamana69aef12011-03-03 15:13:47 +0000267static void filelayout_write_prepare(struct rpc_task *task, void *data)
268{
269 struct nfs_write_data *wdata = (struct nfs_write_data *)data;
270
271 if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
272 &wdata->args.seq_args, &wdata->res.seq_res,
273 0, task))
274 return;
275
276 rpc_call_start(task);
277}
278
279static void filelayout_write_call_done(struct rpc_task *task, void *data)
280{
281 struct nfs_write_data *wdata = (struct nfs_write_data *)data;
282
283 /* Note this may cause RPC to be resent */
284 wdata->mds_ops->rpc_call_done(task, data);
285}
286
287static void filelayout_write_release(void *data)
288{
289 struct nfs_write_data *wdata = (struct nfs_write_data *)data;
290
291 wdata->mds_ops->rpc_release(data);
292}
293
Fred Isamane0c2b382011-03-23 13:27:53 +0000294static void filelayout_commit_release(void *data)
295{
296 struct nfs_write_data *wdata = (struct nfs_write_data *)data;
297
298 nfs_commit_release_pages(wdata);
299 if (atomic_dec_and_test(&NFS_I(wdata->inode)->commits_outstanding))
300 nfs_commit_clear_lock(NFS_I(wdata->inode));
301 nfs_commitdata_release(wdata);
302}
303
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000304struct rpc_call_ops filelayout_read_call_ops = {
305 .rpc_call_prepare = filelayout_read_prepare,
306 .rpc_call_done = filelayout_read_call_done,
307 .rpc_release = filelayout_read_release,
308};
309
Fred Isamana69aef12011-03-03 15:13:47 +0000310struct rpc_call_ops filelayout_write_call_ops = {
311 .rpc_call_prepare = filelayout_write_prepare,
312 .rpc_call_done = filelayout_write_call_done,
313 .rpc_release = filelayout_write_release,
314};
315
Fred Isamane0c2b382011-03-23 13:27:53 +0000316struct rpc_call_ops filelayout_commit_call_ops = {
317 .rpc_call_prepare = filelayout_write_prepare,
318 .rpc_call_done = filelayout_write_call_done,
319 .rpc_release = filelayout_commit_release,
320};
321
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000322static enum pnfs_try_status
323filelayout_read_pagelist(struct nfs_read_data *data)
324{
325 struct pnfs_layout_segment *lseg = data->lseg;
326 struct nfs4_pnfs_ds *ds;
327 loff_t offset = data->args.offset;
328 u32 j, idx;
329 struct nfs_fh *fh;
330 int status;
331
332 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
333 __func__, data->inode->i_ino,
334 data->args.pgbase, (size_t)data->args.count, offset);
335
336 /* Retrieve the correct rpc_client for the byte range */
337 j = nfs4_fl_calc_j_index(lseg, offset);
338 idx = nfs4_fl_calc_ds_index(lseg, j);
339 ds = nfs4_fl_prepare_ds(lseg, idx);
340 if (!ds) {
Andy Adamson568e8c42011-03-01 01:34:22 +0000341 /* Either layout fh index faulty, or ds connect failed */
342 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
343 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000344 return PNFS_NOT_ATTEMPTED;
345 }
346 dprintk("%s USE DS:ip %x %hu\n", __func__,
347 ntohl(ds->ds_ip_addr), ntohs(ds->ds_port));
348
349 /* No multipath support. Use first DS */
350 data->ds_clp = ds->ds_clp;
351 fh = nfs4_fl_select_ds_fh(lseg, j);
352 if (fh)
353 data->args.fh = fh;
354
355 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
356 data->mds_offset = offset;
357
358 /* Perform an asynchronous read to ds */
359 status = nfs_initiate_read(data, ds->ds_clp->cl_rpcclient,
360 &filelayout_read_call_ops);
361 BUG_ON(status != 0);
362 return PNFS_ATTEMPTED;
363}
364
Fred Isamana69aef12011-03-03 15:13:47 +0000365/* Perform async writes. */
Andy Adamson0382b742011-03-03 15:13:45 +0000366static enum pnfs_try_status
367filelayout_write_pagelist(struct nfs_write_data *data, int sync)
368{
Fred Isamana69aef12011-03-03 15:13:47 +0000369 struct pnfs_layout_segment *lseg = data->lseg;
370 struct nfs4_pnfs_ds *ds;
371 loff_t offset = data->args.offset;
372 u32 j, idx;
373 struct nfs_fh *fh;
374 int status;
375
376 /* Retrieve the correct rpc_client for the byte range */
377 j = nfs4_fl_calc_j_index(lseg, offset);
378 idx = nfs4_fl_calc_ds_index(lseg, j);
379 ds = nfs4_fl_prepare_ds(lseg, idx);
380 if (!ds) {
381 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
382 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
383 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
384 return PNFS_NOT_ATTEMPTED;
385 }
386 dprintk("%s ino %lu sync %d req %Zu@%llu DS:%x:%hu\n", __func__,
387 data->inode->i_ino, sync, (size_t) data->args.count, offset,
388 ntohl(ds->ds_ip_addr), ntohs(ds->ds_port));
389
Fred Isamana69aef12011-03-03 15:13:47 +0000390 data->write_done_cb = filelayout_write_done_cb;
391 data->ds_clp = ds->ds_clp;
392 fh = nfs4_fl_select_ds_fh(lseg, j);
393 if (fh)
394 data->args.fh = fh;
395 /*
396 * Get the file offset on the dserver. Set the write offset to
397 * this offset and save the original offset.
398 */
399 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
400 data->mds_offset = offset;
401
402 /* Perform an asynchronous write */
403 status = nfs_initiate_write(data, ds->ds_clp->cl_rpcclient,
404 &filelayout_write_call_ops, sync);
405 BUG_ON(status != 0);
406 return PNFS_ATTEMPTED;
Andy Adamson0382b742011-03-03 15:13:45 +0000407}
408
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000409/*
Andy Adamson16b374c2010-10-20 00:18:04 -0400410 * filelayout_check_layout()
411 *
412 * Make sure layout segment parameters are sane WRT the device.
413 * At this point no generic layer initialization of the lseg has occurred,
414 * and nothing has been added to the layout_hdr cache.
415 *
416 */
417static int
418filelayout_check_layout(struct pnfs_layout_hdr *lo,
419 struct nfs4_filelayout_segment *fl,
420 struct nfs4_layoutget_res *lgr,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400421 struct nfs4_deviceid *id,
422 gfp_t gfp_flags)
Andy Adamson16b374c2010-10-20 00:18:04 -0400423{
Benny Halevya1eaecb2011-05-19 22:14:47 -0400424 struct nfs4_deviceid_node *d;
Andy Adamson16b374c2010-10-20 00:18:04 -0400425 struct nfs4_file_layout_dsaddr *dsaddr;
426 int status = -EINVAL;
Fred Isamanb7edfaa2011-01-06 11:36:21 +0000427 struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
Andy Adamson16b374c2010-10-20 00:18:04 -0400428
429 dprintk("--> %s\n", __func__);
430
431 if (fl->pattern_offset > lgr->range.offset) {
Jim Rees3b6445a2011-02-22 19:31:57 -0500432 dprintk("%s pattern_offset %lld too large\n",
Andy Adamson16b374c2010-10-20 00:18:04 -0400433 __func__, fl->pattern_offset);
434 goto out;
435 }
436
Benny Halevy75247af2011-02-22 15:56:01 -0800437 if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
438 dprintk("%s Invalid stripe unit (%u)\n",
Andy Adamson16b374c2010-10-20 00:18:04 -0400439 __func__, fl->stripe_unit);
440 goto out;
441 }
442
443 /* find and reference the deviceid */
Benny Halevy35c8bb52011-05-24 18:04:02 +0300444 d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
445 NFS_SERVER(lo->plh_inode)->nfs_client, id);
Benny Halevya1eaecb2011-05-19 22:14:47 -0400446 if (d == NULL) {
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400447 dsaddr = get_device_info(lo->plh_inode, id, gfp_flags);
Andy Adamson16b374c2010-10-20 00:18:04 -0400448 if (dsaddr == NULL)
449 goto out;
Benny Halevya1eaecb2011-05-19 22:14:47 -0400450 } else
451 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
Andy Adamson16b374c2010-10-20 00:18:04 -0400452 fl->dsaddr = dsaddr;
453
454 if (fl->first_stripe_index < 0 ||
455 fl->first_stripe_index >= dsaddr->stripe_count) {
456 dprintk("%s Bad first_stripe_index %d\n",
457 __func__, fl->first_stripe_index);
458 goto out_put;
459 }
460
461 if ((fl->stripe_type == STRIPE_SPARSE &&
462 fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
463 (fl->stripe_type == STRIPE_DENSE &&
464 fl->num_fh != dsaddr->stripe_count)) {
465 dprintk("%s num_fh %u not valid for given packing\n",
466 __func__, fl->num_fh);
467 goto out_put;
468 }
469
470 if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
471 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
472 "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
473 nfss->wsize);
474 }
475
476 status = 0;
477out:
478 dprintk("--> %s returns %d\n", __func__, status);
479 return status;
480out_put:
Christoph Hellwigea8eecd2011-03-01 01:34:21 +0000481 nfs4_fl_put_deviceid(dsaddr);
Andy Adamson16b374c2010-10-20 00:18:04 -0400482 goto out;
483}
484
485static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
486{
487 int i;
488
489 for (i = 0; i < fl->num_fh; i++) {
490 if (!fl->fh_array[i])
491 break;
492 kfree(fl->fh_array[i]);
493 }
494 kfree(fl->fh_array);
495 fl->fh_array = NULL;
496}
497
498static void
499_filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
500{
501 filelayout_free_fh_array(fl);
502 kfree(fl);
503}
504
505static int
506filelayout_decode_layout(struct pnfs_layout_hdr *flo,
507 struct nfs4_filelayout_segment *fl,
508 struct nfs4_layoutget_res *lgr,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400509 struct nfs4_deviceid *id,
510 gfp_t gfp_flags)
Andy Adamson16b374c2010-10-20 00:18:04 -0400511{
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400512 struct xdr_stream stream;
Benny Halevyf7da7a12011-05-19 14:16:47 -0400513 struct xdr_buf buf;
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400514 struct page *scratch;
515 __be32 *p;
Andy Adamson16b374c2010-10-20 00:18:04 -0400516 uint32_t nfl_util;
517 int i;
518
519 dprintk("%s: set_layout_map Begin\n", __func__);
520
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400521 scratch = alloc_page(gfp_flags);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400522 if (!scratch)
523 return -ENOMEM;
524
Benny Halevyf7da7a12011-05-19 14:16:47 -0400525 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400526 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
527
528 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
529 * num_fh (4) */
530 p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
531 if (unlikely(!p))
532 goto out_err;
533
Andy Adamson16b374c2010-10-20 00:18:04 -0400534 memcpy(id, p, sizeof(*id));
535 p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
Benny Halevya1eaecb2011-05-19 22:14:47 -0400536 nfs4_print_deviceid(id);
Andy Adamson16b374c2010-10-20 00:18:04 -0400537
538 nfl_util = be32_to_cpup(p++);
539 if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
540 fl->commit_through_mds = 1;
541 if (nfl_util & NFL4_UFLG_DENSE)
542 fl->stripe_type = STRIPE_DENSE;
543 else
544 fl->stripe_type = STRIPE_SPARSE;
545 fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
546
547 fl->first_stripe_index = be32_to_cpup(p++);
548 p = xdr_decode_hyper(p, &fl->pattern_offset);
549 fl->num_fh = be32_to_cpup(p++);
550
551 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
552 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
553 fl->pattern_offset);
554
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400555 if (!fl->num_fh)
556 goto out_err;
557
Andy Adamson16b374c2010-10-20 00:18:04 -0400558 fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *),
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400559 gfp_flags);
Andy Adamson16b374c2010-10-20 00:18:04 -0400560 if (!fl->fh_array)
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400561 goto out_err;
Andy Adamson16b374c2010-10-20 00:18:04 -0400562
563 for (i = 0; i < fl->num_fh; i++) {
564 /* Do we want to use a mempool here? */
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400565 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400566 if (!fl->fh_array[i])
567 goto out_err_free;
568
569 p = xdr_inline_decode(&stream, 4);
570 if (unlikely(!p))
571 goto out_err_free;
Andy Adamson16b374c2010-10-20 00:18:04 -0400572 fl->fh_array[i]->size = be32_to_cpup(p++);
573 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
574 printk(KERN_ERR "Too big fh %d received %d\n",
575 i, fl->fh_array[i]->size);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400576 goto out_err_free;
Andy Adamson16b374c2010-10-20 00:18:04 -0400577 }
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400578
579 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
580 if (unlikely(!p))
581 goto out_err_free;
Andy Adamson16b374c2010-10-20 00:18:04 -0400582 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
Andy Adamson16b374c2010-10-20 00:18:04 -0400583 dprintk("DEBUG: %s: fh len %d\n", __func__,
584 fl->fh_array[i]->size);
585 }
586
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400587 __free_page(scratch);
Andy Adamson16b374c2010-10-20 00:18:04 -0400588 return 0;
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400589
590out_err_free:
591 filelayout_free_fh_array(fl);
592out_err:
593 __free_page(scratch);
594 return -EIO;
Andy Adamson16b374c2010-10-20 00:18:04 -0400595}
596
Fred Isamanc8795132011-03-23 13:27:49 +0000597static void
598filelayout_free_lseg(struct pnfs_layout_segment *lseg)
599{
600 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
601
602 dprintk("--> %s\n", __func__);
603 nfs4_fl_put_deviceid(fl->dsaddr);
Fred Isaman425eb732011-03-23 13:27:50 +0000604 kfree(fl->commit_buckets);
Fred Isamanc8795132011-03-23 13:27:49 +0000605 _filelayout_free_lseg(fl);
606}
607
Andy Adamson16b374c2010-10-20 00:18:04 -0400608static struct pnfs_layout_segment *
609filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400610 struct nfs4_layoutget_res *lgr,
611 gfp_t gfp_flags)
Andy Adamson16b374c2010-10-20 00:18:04 -0400612{
613 struct nfs4_filelayout_segment *fl;
614 int rc;
615 struct nfs4_deviceid id;
616
617 dprintk("--> %s\n", __func__);
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400618 fl = kzalloc(sizeof(*fl), gfp_flags);
Andy Adamson16b374c2010-10-20 00:18:04 -0400619 if (!fl)
620 return NULL;
621
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400622 rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
623 if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
Andy Adamson16b374c2010-10-20 00:18:04 -0400624 _filelayout_free_lseg(fl);
625 return NULL;
626 }
Fred Isaman425eb732011-03-23 13:27:50 +0000627
628 /* This assumes there is only one IOMODE_RW lseg. What
629 * we really want to do is have a layout_hdr level
630 * dictionary of <multipath_list4, fh> keys, each
631 * associated with a struct list_head, populated by calls
632 * to filelayout_write_pagelist().
633 * */
634 if ((!fl->commit_through_mds) && (lgr->range.iomode == IOMODE_RW)) {
635 int i;
636 int size = (fl->stripe_type == STRIPE_SPARSE) ?
637 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
638
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400639 fl->commit_buckets = kcalloc(size, sizeof(struct list_head), gfp_flags);
Fred Isaman425eb732011-03-23 13:27:50 +0000640 if (!fl->commit_buckets) {
641 filelayout_free_lseg(&fl->generic_hdr);
642 return NULL;
643 }
644 fl->number_of_buckets = size;
645 for (i = 0; i < size; i++)
646 INIT_LIST_HEAD(&fl->commit_buckets[i]);
647 }
Andy Adamson16b374c2010-10-20 00:18:04 -0400648 return &fl->generic_hdr;
649}
650
Fred Isaman94ad1c82011-03-01 01:34:14 +0000651/*
652 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
653 *
654 * return 1 : coalesce page
655 * return 0 : don't coalesce page
656 */
657int
658filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
659 struct nfs_page *req)
660{
661 u64 p_stripe, r_stripe;
662 u32 stripe_unit;
663
664 if (!pgio->pg_lseg)
665 return 1;
666 p_stripe = (u64)prev->wb_index << PAGE_CACHE_SHIFT;
667 r_stripe = (u64)req->wb_index << PAGE_CACHE_SHIFT;
668 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
669
670 do_div(p_stripe, stripe_unit);
671 do_div(r_stripe, stripe_unit);
672
673 return (p_stripe == r_stripe);
674}
675
Fred Isamane0c2b382011-03-23 13:27:53 +0000676static bool filelayout_mark_pnfs_commit(struct pnfs_layout_segment *lseg)
677{
678 return !FILELAYOUT_LSEG(lseg)->commit_through_mds;
679}
680
681static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
682{
683 if (fl->stripe_type == STRIPE_SPARSE)
684 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
685 else
686 return j;
687}
688
689struct list_head *filelayout_choose_commit_list(struct nfs_page *req)
690{
691 struct pnfs_layout_segment *lseg = req->wb_commit_lseg;
692 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
693 u32 i, j;
694 struct list_head *list;
695
696 /* Note that we are calling nfs4_fl_calc_j_index on each page
697 * that ends up being committed to a data server. An attractive
698 * alternative is to add a field to nfs_write_data and nfs_page
699 * to store the value calculated in filelayout_write_pagelist
700 * and just use that here.
701 */
702 j = nfs4_fl_calc_j_index(lseg,
703 (loff_t)req->wb_index << PAGE_CACHE_SHIFT);
704 i = select_bucket_index(fl, j);
705 list = &fl->commit_buckets[i];
706 if (list_empty(list)) {
707 /* Non-empty buckets hold a reference on the lseg */
708 get_lseg(lseg);
709 }
710 return list;
711}
712
713static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
714{
715 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
716
717 if (flseg->stripe_type == STRIPE_SPARSE)
718 return i;
719 else
720 return nfs4_fl_calc_ds_index(lseg, i);
721}
722
723static struct nfs_fh *
724select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
725{
726 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
727
728 if (flseg->stripe_type == STRIPE_SPARSE) {
729 if (flseg->num_fh == 1)
730 i = 0;
731 else if (flseg->num_fh == 0)
732 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
733 return NULL;
734 }
735 return flseg->fh_array[i];
736}
737
738static int filelayout_initiate_commit(struct nfs_write_data *data, int how)
739{
740 struct pnfs_layout_segment *lseg = data->lseg;
741 struct nfs4_pnfs_ds *ds;
742 u32 idx;
743 struct nfs_fh *fh;
744
745 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
746 ds = nfs4_fl_prepare_ds(lseg, idx);
747 if (!ds) {
748 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
749 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
750 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
751 prepare_to_resend_writes(data);
752 data->mds_ops->rpc_release(data);
753 return -EAGAIN;
754 }
755 dprintk("%s ino %lu, how %d\n", __func__, data->inode->i_ino, how);
756 data->write_done_cb = filelayout_commit_done_cb;
757 data->ds_clp = ds->ds_clp;
758 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
759 if (fh)
760 data->args.fh = fh;
761 return nfs_initiate_commit(data, ds->ds_clp->cl_rpcclient,
762 &filelayout_commit_call_ops, how);
763}
764
765/*
766 * This is only useful while we are using whole file layouts.
767 */
768static struct pnfs_layout_segment *find_only_write_lseg(struct inode *inode)
769{
770 struct pnfs_layout_segment *lseg, *rv = NULL;
771
772 spin_lock(&inode->i_lock);
773 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
774 if (lseg->pls_range.iomode == IOMODE_RW)
775 rv = get_lseg(lseg);
776 spin_unlock(&inode->i_lock);
777 return rv;
778}
779
780static int alloc_ds_commits(struct inode *inode, struct list_head *list)
781{
782 struct pnfs_layout_segment *lseg;
783 struct nfs4_filelayout_segment *fl;
784 struct nfs_write_data *data;
785 int i, j;
786
787 /* Won't need this when non-whole file layout segments are supported
788 * instead we will use a pnfs_layout_hdr structure */
789 lseg = find_only_write_lseg(inode);
790 if (!lseg)
791 return 0;
792 fl = FILELAYOUT_LSEG(lseg);
793 for (i = 0; i < fl->number_of_buckets; i++) {
794 if (list_empty(&fl->commit_buckets[i]))
795 continue;
796 data = nfs_commitdata_alloc();
797 if (!data)
798 goto out_bad;
799 data->ds_commit_index = i;
800 data->lseg = lseg;
801 list_add(&data->pages, list);
802 }
803 put_lseg(lseg);
804 return 0;
805
806out_bad:
807 for (j = i; j < fl->number_of_buckets; j++) {
808 if (list_empty(&fl->commit_buckets[i]))
809 continue;
810 nfs_retry_commit(&fl->commit_buckets[i], lseg);
811 put_lseg(lseg); /* associated with emptying bucket */
812 }
813 put_lseg(lseg);
814 /* Caller will clean up entries put on list */
815 return -ENOMEM;
816}
817
818/* This follows nfs_commit_list pretty closely */
819static int
820filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
821 int how)
822{
823 struct nfs_write_data *data, *tmp;
824 LIST_HEAD(list);
825
826 if (!list_empty(mds_pages)) {
827 data = nfs_commitdata_alloc();
828 if (!data)
829 goto out_bad;
830 data->lseg = NULL;
831 list_add(&data->pages, &list);
832 }
833
834 if (alloc_ds_commits(inode, &list))
835 goto out_bad;
836
837 list_for_each_entry_safe(data, tmp, &list, pages) {
838 list_del_init(&data->pages);
839 atomic_inc(&NFS_I(inode)->commits_outstanding);
840 if (!data->lseg) {
841 nfs_init_commit(data, mds_pages, NULL);
842 nfs_initiate_commit(data, NFS_CLIENT(inode),
843 data->mds_ops, how);
844 } else {
845 nfs_init_commit(data, &FILELAYOUT_LSEG(data->lseg)->commit_buckets[data->ds_commit_index], data->lseg);
846 filelayout_initiate_commit(data, how);
847 }
848 }
849 return 0;
850 out_bad:
851 list_for_each_entry_safe(data, tmp, &list, pages) {
852 nfs_retry_commit(&data->pages, data->lseg);
853 list_del_init(&data->pages);
854 nfs_commit_free(data);
855 }
856 nfs_retry_commit(mds_pages, NULL);
857 nfs_commit_clear_lock(NFS_I(inode));
858 return -ENOMEM;
859}
860
Benny Halevy1775bc32011-05-20 13:47:33 +0200861static void
862filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
863{
864 nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
865}
866
Dean Hildebrand7ab672c2010-10-20 00:18:00 -0400867static struct pnfs_layoutdriver_type filelayout_type = {
Christoph Hellwigea8eecd2011-03-01 01:34:21 +0000868 .id = LAYOUT_NFSV4_1_FILES,
869 .name = "LAYOUT_NFSV4_1_FILES",
870 .owner = THIS_MODULE,
871 .alloc_lseg = filelayout_alloc_lseg,
872 .free_lseg = filelayout_free_lseg,
Fred Isaman94ad1c82011-03-01 01:34:14 +0000873 .pg_test = filelayout_pg_test,
Fred Isamane0c2b382011-03-23 13:27:53 +0000874 .mark_pnfs_commit = filelayout_mark_pnfs_commit,
875 .choose_commit_list = filelayout_choose_commit_list,
876 .commit_pagelist = filelayout_commit_pagelist,
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000877 .read_pagelist = filelayout_read_pagelist,
Andy Adamson0382b742011-03-03 15:13:45 +0000878 .write_pagelist = filelayout_write_pagelist,
Benny Halevy1775bc32011-05-20 13:47:33 +0200879 .free_deviceid_node = filelayout_free_deveiceid_node,
Dean Hildebrand7ab672c2010-10-20 00:18:00 -0400880};
881
882static int __init nfs4filelayout_init(void)
883{
884 printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
885 __func__);
886 return pnfs_register_layoutdriver(&filelayout_type);
887}
888
889static void __exit nfs4filelayout_exit(void)
890{
891 printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
892 __func__);
893 pnfs_unregister_layoutdriver(&filelayout_type);
894}
895
896module_init(nfs4filelayout_init);
897module_exit(nfs4filelayout_exit);