blob: 0bafcc91c27f8d8513dc1a7c8776fefd3411eee2 [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>
Andy Adamson16b374c2010-10-20 00:18:04 -040034
35#include "internal.h"
36#include "nfs4filelayout.h"
Dean Hildebrand7ab672c2010-10-20 00:18:00 -040037
38#define NFSDBG_FACILITY NFSDBG_PNFS_LD
39
40MODULE_LICENSE("GPL");
41MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
42MODULE_DESCRIPTION("The NFSv4 file layout driver");
43
Andy Adamsoncbdabc72011-03-01 01:34:20 +000044#define FILELAYOUT_POLL_RETRY_MAX (15*HZ)
45
Fred Isamancfe7f412011-03-01 01:34:18 +000046static loff_t
47filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
48 loff_t offset)
49{
50 u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
51 u64 tmp;
52
53 offset -= flseg->pattern_offset;
54 tmp = offset;
55 do_div(tmp, stripe_width);
56
57 return tmp * flseg->stripe_unit + do_div(offset, flseg->stripe_unit);
58}
59
60/* This function is used by the layout driver to calculate the
61 * offset of the file on the dserver based on whether the
62 * layout type is STRIPE_DENSE or STRIPE_SPARSE
63 */
64static loff_t
65filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
66{
67 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
68
69 switch (flseg->stripe_type) {
70 case STRIPE_SPARSE:
71 return offset;
72
73 case STRIPE_DENSE:
74 return filelayout_get_dense_offset(flseg, offset);
75 }
76
77 BUG();
78}
79
Andy Adamsoncbdabc72011-03-01 01:34:20 +000080/* For data server errors we don't recover from */
81static void
82filelayout_set_lo_fail(struct pnfs_layout_segment *lseg)
83{
84 if (lseg->pls_range.iomode == IOMODE_RW) {
85 dprintk("%s Setting layout IOMODE_RW fail bit\n", __func__);
86 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
87 } else {
88 dprintk("%s Setting layout IOMODE_READ fail bit\n", __func__);
89 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
90 }
91}
92
93static int filelayout_async_handle_error(struct rpc_task *task,
94 struct nfs4_state *state,
95 struct nfs_client *clp,
96 int *reset)
97{
98 if (task->tk_status >= 0)
99 return 0;
100
101 *reset = 0;
102
103 switch (task->tk_status) {
104 case -NFS4ERR_BADSESSION:
105 case -NFS4ERR_BADSLOT:
106 case -NFS4ERR_BAD_HIGH_SLOT:
107 case -NFS4ERR_DEADSESSION:
108 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
109 case -NFS4ERR_SEQ_FALSE_RETRY:
110 case -NFS4ERR_SEQ_MISORDERED:
111 dprintk("%s ERROR %d, Reset session. Exchangeid "
112 "flags 0x%x\n", __func__, task->tk_status,
113 clp->cl_exchange_flags);
114 nfs4_schedule_session_recovery(clp->cl_session);
115 break;
116 case -NFS4ERR_DELAY:
117 case -NFS4ERR_GRACE:
118 case -EKEYEXPIRED:
119 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
120 break;
Andy Adamsona8a4ae32011-05-03 13:43:03 -0400121 case -NFS4ERR_RETRY_UNCACHED_REP:
122 break;
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000123 default:
124 dprintk("%s DS error. Retry through MDS %d\n", __func__,
125 task->tk_status);
126 *reset = 1;
127 break;
128 }
129 task->tk_status = 0;
130 return -EAGAIN;
131}
132
133/* NFS_PROTO call done callback routines */
134
135static int filelayout_read_done_cb(struct rpc_task *task,
136 struct nfs_read_data *data)
137{
138 struct nfs_client *clp = data->ds_clp;
139 int reset = 0;
140
141 dprintk("%s DS read\n", __func__);
142
143 if (filelayout_async_handle_error(task, data->args.context->state,
144 data->ds_clp, &reset) == -EAGAIN) {
145 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
146 __func__, data->ds_clp, data->ds_clp->cl_session);
147 if (reset) {
148 filelayout_set_lo_fail(data->lseg);
149 nfs4_reset_read(task, data);
150 clp = NFS_SERVER(data->inode)->nfs_client;
151 }
152 nfs_restart_rpc(task, clp);
153 return -EAGAIN;
154 }
155
156 return 0;
157}
158
Andy Adamson16b374c2010-10-20 00:18:04 -0400159/*
Andy Adamson863a3c62011-03-23 13:27:54 +0000160 * We reference the rpc_cred of the first WRITE that triggers the need for
161 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
162 * rfc5661 is not clear about which credential should be used.
163 */
164static void
165filelayout_set_layoutcommit(struct nfs_write_data *wdata)
166{
167 if (FILELAYOUT_LSEG(wdata->lseg)->commit_through_mds ||
168 wdata->res.verf->committed == NFS_FILE_SYNC)
169 return;
170
171 pnfs_set_layoutcommit(wdata);
172 dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, wdata->inode->i_ino,
173 (unsigned long) wdata->lseg->pls_end_pos);
174}
175
176/*
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000177 * Call ops for the async read/write cases
178 * In the case of dense layouts, the offset needs to be reset to its
179 * original value.
180 */
181static void filelayout_read_prepare(struct rpc_task *task, void *data)
182{
183 struct nfs_read_data *rdata = (struct nfs_read_data *)data;
184
Andy Adamsoncbdabc72011-03-01 01:34:20 +0000185 rdata->read_done_cb = filelayout_read_done_cb;
186
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000187 if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
188 &rdata->args.seq_args, &rdata->res.seq_res,
189 0, task))
190 return;
191
192 rpc_call_start(task);
193}
194
195static void filelayout_read_call_done(struct rpc_task *task, void *data)
196{
197 struct nfs_read_data *rdata = (struct nfs_read_data *)data;
198
199 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
200
201 /* Note this may cause RPC to be resent */
202 rdata->mds_ops->rpc_call_done(task, data);
203}
204
205static void filelayout_read_release(void *data)
206{
207 struct nfs_read_data *rdata = (struct nfs_read_data *)data;
208
209 rdata->mds_ops->rpc_release(data);
210}
211
Fred Isamana69aef12011-03-03 15:13:47 +0000212static int filelayout_write_done_cb(struct rpc_task *task,
213 struct nfs_write_data *data)
214{
215 int reset = 0;
216
217 if (filelayout_async_handle_error(task, data->args.context->state,
218 data->ds_clp, &reset) == -EAGAIN) {
219 struct nfs_client *clp;
220
221 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
222 __func__, data->ds_clp, data->ds_clp->cl_session);
223 if (reset) {
224 filelayout_set_lo_fail(data->lseg);
225 nfs4_reset_write(task, data);
226 clp = NFS_SERVER(data->inode)->nfs_client;
227 } else
228 clp = data->ds_clp;
229 nfs_restart_rpc(task, clp);
230 return -EAGAIN;
231 }
232
Andy Adamson863a3c62011-03-23 13:27:54 +0000233 filelayout_set_layoutcommit(data);
Fred Isamana69aef12011-03-03 15:13:47 +0000234 return 0;
235}
236
Fred Isamane0c2b382011-03-23 13:27:53 +0000237/* Fake up some data that will cause nfs_commit_release to retry the writes. */
238static void prepare_to_resend_writes(struct nfs_write_data *data)
239{
240 struct nfs_page *first = nfs_list_entry(data->pages.next);
241
242 data->task.tk_status = 0;
243 memcpy(data->verf.verifier, first->wb_verf.verifier,
244 sizeof(first->wb_verf.verifier));
245 data->verf.verifier[0]++; /* ensure verifier mismatch */
246}
247
248static int filelayout_commit_done_cb(struct rpc_task *task,
249 struct nfs_write_data *data)
250{
251 int reset = 0;
252
253 if (filelayout_async_handle_error(task, data->args.context->state,
254 data->ds_clp, &reset) == -EAGAIN) {
255 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
256 __func__, data->ds_clp, data->ds_clp->cl_session);
257 if (reset) {
258 prepare_to_resend_writes(data);
259 filelayout_set_lo_fail(data->lseg);
260 } else
261 nfs_restart_rpc(task, data->ds_clp);
262 return -EAGAIN;
263 }
264
265 return 0;
266}
267
Fred Isamana69aef12011-03-03 15:13:47 +0000268static void filelayout_write_prepare(struct rpc_task *task, void *data)
269{
270 struct nfs_write_data *wdata = (struct nfs_write_data *)data;
271
272 if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
273 &wdata->args.seq_args, &wdata->res.seq_res,
274 0, task))
275 return;
276
277 rpc_call_start(task);
278}
279
280static void filelayout_write_call_done(struct rpc_task *task, void *data)
281{
282 struct nfs_write_data *wdata = (struct nfs_write_data *)data;
283
284 /* Note this may cause RPC to be resent */
285 wdata->mds_ops->rpc_call_done(task, data);
286}
287
288static void filelayout_write_release(void *data)
289{
290 struct nfs_write_data *wdata = (struct nfs_write_data *)data;
291
292 wdata->mds_ops->rpc_release(data);
293}
294
Fred Isamane0c2b382011-03-23 13:27:53 +0000295static void filelayout_commit_release(void *data)
296{
297 struct nfs_write_data *wdata = (struct nfs_write_data *)data;
298
299 nfs_commit_release_pages(wdata);
300 if (atomic_dec_and_test(&NFS_I(wdata->inode)->commits_outstanding))
301 nfs_commit_clear_lock(NFS_I(wdata->inode));
302 nfs_commitdata_release(wdata);
303}
304
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000305struct rpc_call_ops filelayout_read_call_ops = {
306 .rpc_call_prepare = filelayout_read_prepare,
307 .rpc_call_done = filelayout_read_call_done,
308 .rpc_release = filelayout_read_release,
309};
310
Fred Isamana69aef12011-03-03 15:13:47 +0000311struct rpc_call_ops filelayout_write_call_ops = {
312 .rpc_call_prepare = filelayout_write_prepare,
313 .rpc_call_done = filelayout_write_call_done,
314 .rpc_release = filelayout_write_release,
315};
316
Fred Isamane0c2b382011-03-23 13:27:53 +0000317struct rpc_call_ops filelayout_commit_call_ops = {
318 .rpc_call_prepare = filelayout_write_prepare,
319 .rpc_call_done = filelayout_write_call_done,
320 .rpc_release = filelayout_commit_release,
321};
322
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000323static enum pnfs_try_status
324filelayout_read_pagelist(struct nfs_read_data *data)
325{
326 struct pnfs_layout_segment *lseg = data->lseg;
327 struct nfs4_pnfs_ds *ds;
328 loff_t offset = data->args.offset;
329 u32 j, idx;
330 struct nfs_fh *fh;
331 int status;
332
333 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
334 __func__, data->inode->i_ino,
335 data->args.pgbase, (size_t)data->args.count, offset);
336
337 /* Retrieve the correct rpc_client for the byte range */
338 j = nfs4_fl_calc_j_index(lseg, offset);
339 idx = nfs4_fl_calc_ds_index(lseg, j);
340 ds = nfs4_fl_prepare_ds(lseg, idx);
341 if (!ds) {
Andy Adamson568e8c42011-03-01 01:34:22 +0000342 /* Either layout fh index faulty, or ds connect failed */
343 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
344 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000345 return PNFS_NOT_ATTEMPTED;
346 }
347 dprintk("%s USE DS:ip %x %hu\n", __func__,
348 ntohl(ds->ds_ip_addr), ntohs(ds->ds_port));
349
350 /* No multipath support. Use first DS */
351 data->ds_clp = ds->ds_clp;
352 fh = nfs4_fl_select_ds_fh(lseg, j);
353 if (fh)
354 data->args.fh = fh;
355
356 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
357 data->mds_offset = offset;
358
359 /* Perform an asynchronous read to ds */
360 status = nfs_initiate_read(data, ds->ds_clp->cl_rpcclient,
361 &filelayout_read_call_ops);
362 BUG_ON(status != 0);
363 return PNFS_ATTEMPTED;
364}
365
Fred Isamana69aef12011-03-03 15:13:47 +0000366/* Perform async writes. */
Andy Adamson0382b742011-03-03 15:13:45 +0000367static enum pnfs_try_status
368filelayout_write_pagelist(struct nfs_write_data *data, int sync)
369{
Fred Isamana69aef12011-03-03 15:13:47 +0000370 struct pnfs_layout_segment *lseg = data->lseg;
371 struct nfs4_pnfs_ds *ds;
372 loff_t offset = data->args.offset;
373 u32 j, idx;
374 struct nfs_fh *fh;
375 int status;
376
377 /* Retrieve the correct rpc_client for the byte range */
378 j = nfs4_fl_calc_j_index(lseg, offset);
379 idx = nfs4_fl_calc_ds_index(lseg, j);
380 ds = nfs4_fl_prepare_ds(lseg, idx);
381 if (!ds) {
382 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
383 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
384 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
385 return PNFS_NOT_ATTEMPTED;
386 }
387 dprintk("%s ino %lu sync %d req %Zu@%llu DS:%x:%hu\n", __func__,
388 data->inode->i_ino, sync, (size_t) data->args.count, offset,
389 ntohl(ds->ds_ip_addr), ntohs(ds->ds_port));
390
Fred Isamana69aef12011-03-03 15:13:47 +0000391 data->write_done_cb = filelayout_write_done_cb;
392 data->ds_clp = ds->ds_clp;
393 fh = nfs4_fl_select_ds_fh(lseg, j);
394 if (fh)
395 data->args.fh = fh;
396 /*
397 * Get the file offset on the dserver. Set the write offset to
398 * this offset and save the original offset.
399 */
400 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
401 data->mds_offset = offset;
402
403 /* Perform an asynchronous write */
404 status = nfs_initiate_write(data, ds->ds_clp->cl_rpcclient,
405 &filelayout_write_call_ops, sync);
406 BUG_ON(status != 0);
407 return PNFS_ATTEMPTED;
Andy Adamson0382b742011-03-03 15:13:45 +0000408}
409
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000410/*
Andy Adamson16b374c2010-10-20 00:18:04 -0400411 * filelayout_check_layout()
412 *
413 * Make sure layout segment parameters are sane WRT the device.
414 * At this point no generic layer initialization of the lseg has occurred,
415 * and nothing has been added to the layout_hdr cache.
416 *
417 */
418static int
419filelayout_check_layout(struct pnfs_layout_hdr *lo,
420 struct nfs4_filelayout_segment *fl,
421 struct nfs4_layoutget_res *lgr,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400422 struct nfs4_deviceid *id,
423 gfp_t gfp_flags)
Andy Adamson16b374c2010-10-20 00:18:04 -0400424{
Benny Halevya1eaecb2011-05-19 22:14:47 -0400425 struct nfs4_deviceid_node *d;
Andy Adamson16b374c2010-10-20 00:18:04 -0400426 struct nfs4_file_layout_dsaddr *dsaddr;
427 int status = -EINVAL;
Fred Isamanb7edfaa2011-01-06 11:36:21 +0000428 struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
Andy Adamson16b374c2010-10-20 00:18:04 -0400429
430 dprintk("--> %s\n", __func__);
431
432 if (fl->pattern_offset > lgr->range.offset) {
Jim Rees3b6445a2011-02-22 19:31:57 -0500433 dprintk("%s pattern_offset %lld too large\n",
Andy Adamson16b374c2010-10-20 00:18:04 -0400434 __func__, fl->pattern_offset);
435 goto out;
436 }
437
Benny Halevy75247af2011-02-22 15:56:01 -0800438 if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
439 dprintk("%s Invalid stripe unit (%u)\n",
Andy Adamson16b374c2010-10-20 00:18:04 -0400440 __func__, fl->stripe_unit);
441 goto out;
442 }
443
444 /* find and reference the deviceid */
Benny Halevy35c8bb52011-05-24 18:04:02 +0300445 d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
446 NFS_SERVER(lo->plh_inode)->nfs_client, id);
Benny Halevya1eaecb2011-05-19 22:14:47 -0400447 if (d == NULL) {
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400448 dsaddr = get_device_info(lo->plh_inode, id, gfp_flags);
Andy Adamson16b374c2010-10-20 00:18:04 -0400449 if (dsaddr == NULL)
450 goto out;
Benny Halevya1eaecb2011-05-19 22:14:47 -0400451 } else
452 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
Andy Adamson16b374c2010-10-20 00:18:04 -0400453 fl->dsaddr = dsaddr;
454
455 if (fl->first_stripe_index < 0 ||
456 fl->first_stripe_index >= dsaddr->stripe_count) {
457 dprintk("%s Bad first_stripe_index %d\n",
458 __func__, fl->first_stripe_index);
459 goto out_put;
460 }
461
462 if ((fl->stripe_type == STRIPE_SPARSE &&
463 fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
464 (fl->stripe_type == STRIPE_DENSE &&
465 fl->num_fh != dsaddr->stripe_count)) {
466 dprintk("%s num_fh %u not valid for given packing\n",
467 __func__, fl->num_fh);
468 goto out_put;
469 }
470
471 if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
472 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
473 "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
474 nfss->wsize);
475 }
476
477 status = 0;
478out:
479 dprintk("--> %s returns %d\n", __func__, status);
480 return status;
481out_put:
Christoph Hellwigea8eecd2011-03-01 01:34:21 +0000482 nfs4_fl_put_deviceid(dsaddr);
Andy Adamson16b374c2010-10-20 00:18:04 -0400483 goto out;
484}
485
486static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
487{
488 int i;
489
490 for (i = 0; i < fl->num_fh; i++) {
491 if (!fl->fh_array[i])
492 break;
493 kfree(fl->fh_array[i]);
494 }
495 kfree(fl->fh_array);
496 fl->fh_array = NULL;
497}
498
499static void
500_filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
501{
502 filelayout_free_fh_array(fl);
503 kfree(fl);
504}
505
506static int
507filelayout_decode_layout(struct pnfs_layout_hdr *flo,
508 struct nfs4_filelayout_segment *fl,
509 struct nfs4_layoutget_res *lgr,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400510 struct nfs4_deviceid *id,
511 gfp_t gfp_flags)
Andy Adamson16b374c2010-10-20 00:18:04 -0400512{
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400513 struct xdr_stream stream;
Benny Halevyf7da7a12011-05-19 14:16:47 -0400514 struct xdr_buf buf;
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400515 struct page *scratch;
516 __be32 *p;
Andy Adamson16b374c2010-10-20 00:18:04 -0400517 uint32_t nfl_util;
518 int i;
519
520 dprintk("%s: set_layout_map Begin\n", __func__);
521
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400522 scratch = alloc_page(gfp_flags);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400523 if (!scratch)
524 return -ENOMEM;
525
Benny Halevyf7da7a12011-05-19 14:16:47 -0400526 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400527 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
528
529 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
530 * num_fh (4) */
531 p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
532 if (unlikely(!p))
533 goto out_err;
534
Andy Adamson16b374c2010-10-20 00:18:04 -0400535 memcpy(id, p, sizeof(*id));
536 p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
Benny Halevya1eaecb2011-05-19 22:14:47 -0400537 nfs4_print_deviceid(id);
Andy Adamson16b374c2010-10-20 00:18:04 -0400538
539 nfl_util = be32_to_cpup(p++);
540 if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
541 fl->commit_through_mds = 1;
542 if (nfl_util & NFL4_UFLG_DENSE)
543 fl->stripe_type = STRIPE_DENSE;
544 else
545 fl->stripe_type = STRIPE_SPARSE;
546 fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
547
548 fl->first_stripe_index = be32_to_cpup(p++);
549 p = xdr_decode_hyper(p, &fl->pattern_offset);
550 fl->num_fh = be32_to_cpup(p++);
551
552 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
553 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
554 fl->pattern_offset);
555
Andy Adamsoncec765c2011-06-13 18:36:17 -0400556 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
557 * Futher checking is done in filelayout_check_layout */
558 if (fl->num_fh < 0 || fl->num_fh >
559 max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400560 goto out_err;
561
Andy Adamsoncec765c2011-06-13 18:36:17 -0400562 if (fl->num_fh > 0) {
563 fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *),
564 gfp_flags);
565 if (!fl->fh_array)
566 goto out_err;
567 }
Andy Adamson16b374c2010-10-20 00:18:04 -0400568
569 for (i = 0; i < fl->num_fh; i++) {
570 /* Do we want to use a mempool here? */
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400571 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400572 if (!fl->fh_array[i])
573 goto out_err_free;
574
575 p = xdr_inline_decode(&stream, 4);
576 if (unlikely(!p))
577 goto out_err_free;
Andy Adamson16b374c2010-10-20 00:18:04 -0400578 fl->fh_array[i]->size = be32_to_cpup(p++);
579 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
580 printk(KERN_ERR "Too big fh %d received %d\n",
581 i, fl->fh_array[i]->size);
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400582 goto out_err_free;
Andy Adamson16b374c2010-10-20 00:18:04 -0400583 }
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400584
585 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
586 if (unlikely(!p))
587 goto out_err_free;
Andy Adamson16b374c2010-10-20 00:18:04 -0400588 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
Andy Adamson16b374c2010-10-20 00:18:04 -0400589 dprintk("DEBUG: %s: fh len %d\n", __func__,
590 fl->fh_array[i]->size);
591 }
592
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400593 __free_page(scratch);
Andy Adamson16b374c2010-10-20 00:18:04 -0400594 return 0;
Weston Andros Adamson35124a02011-03-24 16:48:21 -0400595
596out_err_free:
597 filelayout_free_fh_array(fl);
598out_err:
599 __free_page(scratch);
600 return -EIO;
Andy Adamson16b374c2010-10-20 00:18:04 -0400601}
602
Fred Isamanc8795132011-03-23 13:27:49 +0000603static void
604filelayout_free_lseg(struct pnfs_layout_segment *lseg)
605{
606 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
607
608 dprintk("--> %s\n", __func__);
609 nfs4_fl_put_deviceid(fl->dsaddr);
Fred Isaman425eb732011-03-23 13:27:50 +0000610 kfree(fl->commit_buckets);
Fred Isamanc8795132011-03-23 13:27:49 +0000611 _filelayout_free_lseg(fl);
612}
613
Andy Adamson16b374c2010-10-20 00:18:04 -0400614static struct pnfs_layout_segment *
615filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400616 struct nfs4_layoutget_res *lgr,
617 gfp_t gfp_flags)
Andy Adamson16b374c2010-10-20 00:18:04 -0400618{
619 struct nfs4_filelayout_segment *fl;
620 int rc;
621 struct nfs4_deviceid id;
622
623 dprintk("--> %s\n", __func__);
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400624 fl = kzalloc(sizeof(*fl), gfp_flags);
Andy Adamson16b374c2010-10-20 00:18:04 -0400625 if (!fl)
626 return NULL;
627
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400628 rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
629 if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
Andy Adamson16b374c2010-10-20 00:18:04 -0400630 _filelayout_free_lseg(fl);
631 return NULL;
632 }
Fred Isaman425eb732011-03-23 13:27:50 +0000633
634 /* This assumes there is only one IOMODE_RW lseg. What
635 * we really want to do is have a layout_hdr level
636 * dictionary of <multipath_list4, fh> keys, each
637 * associated with a struct list_head, populated by calls
638 * to filelayout_write_pagelist().
639 * */
640 if ((!fl->commit_through_mds) && (lgr->range.iomode == IOMODE_RW)) {
641 int i;
642 int size = (fl->stripe_type == STRIPE_SPARSE) ?
643 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
644
Trond Myklebusta75b9df2011-05-11 18:00:51 -0400645 fl->commit_buckets = kcalloc(size, sizeof(struct list_head), gfp_flags);
Fred Isaman425eb732011-03-23 13:27:50 +0000646 if (!fl->commit_buckets) {
647 filelayout_free_lseg(&fl->generic_hdr);
648 return NULL;
649 }
650 fl->number_of_buckets = size;
651 for (i = 0; i < size; i++)
652 INIT_LIST_HEAD(&fl->commit_buckets[i]);
653 }
Andy Adamson16b374c2010-10-20 00:18:04 -0400654 return &fl->generic_hdr;
655}
656
Fred Isaman94ad1c82011-03-01 01:34:14 +0000657/*
658 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
659 *
Benny Halevy18ad0a92011-05-25 21:03:56 +0300660 * return true : coalesce page
661 * return false : don't coalesce page
Fred Isaman94ad1c82011-03-01 01:34:14 +0000662 */
Benny Halevy18ad0a92011-05-25 21:03:56 +0300663bool
Fred Isaman94ad1c82011-03-01 01:34:14 +0000664filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
665 struct nfs_page *req)
666{
667 u64 p_stripe, r_stripe;
668 u32 stripe_unit;
669
Benny Halevy19345cb2011-06-19 18:33:46 -0400670 if (!pnfs_generic_pg_test(pgio, prev, req) ||
671 !nfs_generic_pg_test(pgio, prev, req))
672 return false;
Benny Halevy89a58e32011-05-25 20:54:40 +0300673
Fred Isaman94ad1c82011-03-01 01:34:14 +0000674 if (!pgio->pg_lseg)
675 return 1;
676 p_stripe = (u64)prev->wb_index << PAGE_CACHE_SHIFT;
677 r_stripe = (u64)req->wb_index << PAGE_CACHE_SHIFT;
678 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
679
680 do_div(p_stripe, stripe_unit);
681 do_div(r_stripe, stripe_unit);
682
683 return (p_stripe == r_stripe);
684}
685
Fred Isamane0c2b382011-03-23 13:27:53 +0000686static bool filelayout_mark_pnfs_commit(struct pnfs_layout_segment *lseg)
687{
688 return !FILELAYOUT_LSEG(lseg)->commit_through_mds;
689}
690
691static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
692{
693 if (fl->stripe_type == STRIPE_SPARSE)
694 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
695 else
696 return j;
697}
698
699struct list_head *filelayout_choose_commit_list(struct nfs_page *req)
700{
701 struct pnfs_layout_segment *lseg = req->wb_commit_lseg;
702 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
703 u32 i, j;
704 struct list_head *list;
705
706 /* Note that we are calling nfs4_fl_calc_j_index on each page
707 * that ends up being committed to a data server. An attractive
708 * alternative is to add a field to nfs_write_data and nfs_page
709 * to store the value calculated in filelayout_write_pagelist
710 * and just use that here.
711 */
712 j = nfs4_fl_calc_j_index(lseg,
713 (loff_t)req->wb_index << PAGE_CACHE_SHIFT);
714 i = select_bucket_index(fl, j);
715 list = &fl->commit_buckets[i];
716 if (list_empty(list)) {
717 /* Non-empty buckets hold a reference on the lseg */
718 get_lseg(lseg);
719 }
720 return list;
721}
722
723static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
724{
725 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
726
727 if (flseg->stripe_type == STRIPE_SPARSE)
728 return i;
729 else
730 return nfs4_fl_calc_ds_index(lseg, i);
731}
732
733static struct nfs_fh *
734select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
735{
736 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
737
738 if (flseg->stripe_type == STRIPE_SPARSE) {
739 if (flseg->num_fh == 1)
740 i = 0;
741 else if (flseg->num_fh == 0)
742 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
743 return NULL;
744 }
745 return flseg->fh_array[i];
746}
747
748static int filelayout_initiate_commit(struct nfs_write_data *data, int how)
749{
750 struct pnfs_layout_segment *lseg = data->lseg;
751 struct nfs4_pnfs_ds *ds;
752 u32 idx;
753 struct nfs_fh *fh;
754
755 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
756 ds = nfs4_fl_prepare_ds(lseg, idx);
757 if (!ds) {
758 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
759 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
760 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
761 prepare_to_resend_writes(data);
762 data->mds_ops->rpc_release(data);
763 return -EAGAIN;
764 }
765 dprintk("%s ino %lu, how %d\n", __func__, data->inode->i_ino, how);
766 data->write_done_cb = filelayout_commit_done_cb;
767 data->ds_clp = ds->ds_clp;
768 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
769 if (fh)
770 data->args.fh = fh;
771 return nfs_initiate_commit(data, ds->ds_clp->cl_rpcclient,
772 &filelayout_commit_call_ops, how);
773}
774
775/*
776 * This is only useful while we are using whole file layouts.
777 */
778static struct pnfs_layout_segment *find_only_write_lseg(struct inode *inode)
779{
780 struct pnfs_layout_segment *lseg, *rv = NULL;
781
782 spin_lock(&inode->i_lock);
783 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
784 if (lseg->pls_range.iomode == IOMODE_RW)
785 rv = get_lseg(lseg);
786 spin_unlock(&inode->i_lock);
787 return rv;
788}
789
790static int alloc_ds_commits(struct inode *inode, struct list_head *list)
791{
792 struct pnfs_layout_segment *lseg;
793 struct nfs4_filelayout_segment *fl;
794 struct nfs_write_data *data;
795 int i, j;
796
797 /* Won't need this when non-whole file layout segments are supported
798 * instead we will use a pnfs_layout_hdr structure */
799 lseg = find_only_write_lseg(inode);
800 if (!lseg)
801 return 0;
802 fl = FILELAYOUT_LSEG(lseg);
803 for (i = 0; i < fl->number_of_buckets; i++) {
804 if (list_empty(&fl->commit_buckets[i]))
805 continue;
806 data = nfs_commitdata_alloc();
807 if (!data)
808 goto out_bad;
809 data->ds_commit_index = i;
810 data->lseg = lseg;
811 list_add(&data->pages, list);
812 }
813 put_lseg(lseg);
814 return 0;
815
816out_bad:
817 for (j = i; j < fl->number_of_buckets; j++) {
818 if (list_empty(&fl->commit_buckets[i]))
819 continue;
820 nfs_retry_commit(&fl->commit_buckets[i], lseg);
821 put_lseg(lseg); /* associated with emptying bucket */
822 }
823 put_lseg(lseg);
824 /* Caller will clean up entries put on list */
825 return -ENOMEM;
826}
827
828/* This follows nfs_commit_list pretty closely */
829static int
830filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
831 int how)
832{
833 struct nfs_write_data *data, *tmp;
834 LIST_HEAD(list);
835
836 if (!list_empty(mds_pages)) {
837 data = nfs_commitdata_alloc();
838 if (!data)
839 goto out_bad;
840 data->lseg = NULL;
841 list_add(&data->pages, &list);
842 }
843
844 if (alloc_ds_commits(inode, &list))
845 goto out_bad;
846
847 list_for_each_entry_safe(data, tmp, &list, pages) {
848 list_del_init(&data->pages);
849 atomic_inc(&NFS_I(inode)->commits_outstanding);
850 if (!data->lseg) {
851 nfs_init_commit(data, mds_pages, NULL);
852 nfs_initiate_commit(data, NFS_CLIENT(inode),
853 data->mds_ops, how);
854 } else {
855 nfs_init_commit(data, &FILELAYOUT_LSEG(data->lseg)->commit_buckets[data->ds_commit_index], data->lseg);
856 filelayout_initiate_commit(data, how);
857 }
858 }
859 return 0;
860 out_bad:
861 list_for_each_entry_safe(data, tmp, &list, pages) {
862 nfs_retry_commit(&data->pages, data->lseg);
863 list_del_init(&data->pages);
864 nfs_commit_free(data);
865 }
866 nfs_retry_commit(mds_pages, NULL);
867 nfs_commit_clear_lock(NFS_I(inode));
868 return -ENOMEM;
869}
870
Benny Halevy1775bc32011-05-20 13:47:33 +0200871static void
872filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
873{
874 nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
875}
876
Dean Hildebrand7ab672c2010-10-20 00:18:00 -0400877static struct pnfs_layoutdriver_type filelayout_type = {
Christoph Hellwigea8eecd2011-03-01 01:34:21 +0000878 .id = LAYOUT_NFSV4_1_FILES,
879 .name = "LAYOUT_NFSV4_1_FILES",
880 .owner = THIS_MODULE,
881 .alloc_lseg = filelayout_alloc_lseg,
882 .free_lseg = filelayout_free_lseg,
Fred Isaman94ad1c82011-03-01 01:34:14 +0000883 .pg_test = filelayout_pg_test,
Fred Isamane0c2b382011-03-23 13:27:53 +0000884 .mark_pnfs_commit = filelayout_mark_pnfs_commit,
885 .choose_commit_list = filelayout_choose_commit_list,
886 .commit_pagelist = filelayout_commit_pagelist,
Andy Adamsondc70d7b2011-03-01 01:34:19 +0000887 .read_pagelist = filelayout_read_pagelist,
Andy Adamson0382b742011-03-03 15:13:45 +0000888 .write_pagelist = filelayout_write_pagelist,
Benny Halevy1775bc32011-05-20 13:47:33 +0200889 .free_deviceid_node = filelayout_free_deveiceid_node,
Dean Hildebrand7ab672c2010-10-20 00:18:00 -0400890};
891
892static int __init nfs4filelayout_init(void)
893{
894 printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
895 __func__);
896 return pnfs_register_layoutdriver(&filelayout_type);
897}
898
899static void __exit nfs4filelayout_exit(void)
900{
901 printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
902 __func__);
903 pnfs_unregister_layoutdriver(&filelayout_type);
904}
905
906module_init(nfs4filelayout_init);
907module_exit(nfs4filelayout_exit);