Commit | Line | Data |
---|---|---|
85e174ba RL |
1 | /* |
2 | * pNFS functions to call and manage layout drivers. | |
3 | * | |
4 | * Copyright (c) 2002 [year of first publication] | |
5 | * The Regents of the University of Michigan | |
6 | * All Rights Reserved | |
7 | * | |
8 | * Dean Hildebrand <dhildebz@umich.edu> | |
9 | * | |
10 | * Permission is granted to use, copy, create derivative works, and | |
11 | * redistribute this software and such derivative works for any purpose, | |
12 | * so long as the name of the University of Michigan is not used in | |
13 | * any advertising or publicity pertaining to the use or distribution | |
14 | * of this software without specific, written prior authorization. If | |
15 | * the above copyright notice or any other identification of the | |
16 | * University of Michigan is included in any copy of any portion of | |
17 | * this software, then the disclaimer below must also be included. | |
18 | * | |
19 | * This software is provided as is, without representation or warranty | |
20 | * of any kind either express or implied, including without limitation | |
21 | * the implied warranties of merchantability, fitness for a particular | |
22 | * purpose, or noninfringement. The Regents of the University of | |
23 | * Michigan shall not be liable for any damages, including special, | |
24 | * indirect, incidental, or consequential damages, with respect to any | |
25 | * claim arising out of or in connection with the use of the software, | |
26 | * even if it has been or is hereafter advised of the possibility of | |
27 | * such damages. | |
28 | */ | |
29 | ||
30 | #include <linux/nfs_fs.h> | |
493292dd | 31 | #include <linux/nfs_page.h> |
143cb494 | 32 | #include <linux/module.h> |
974cec8c | 33 | #include "internal.h" |
85e174ba | 34 | #include "pnfs.h" |
64419a9b | 35 | #include "iostat.h" |
cc668ab3 | 36 | #include "nfs4trace.h" |
40dd4b7a | 37 | #include "delegation.h" |
8733408d | 38 | #include "nfs42.h" |
85e174ba RL |
39 | |
40 | #define NFSDBG_FACILITY NFSDBG_PNFS | |
25c75333 | 41 | #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ) |
85e174ba | 42 | |
02c35fca FI |
43 | /* Locking: |
44 | * | |
45 | * pnfs_spinlock: | |
46 | * protects pnfs_modules_tbl. | |
47 | */ | |
48 | static DEFINE_SPINLOCK(pnfs_spinlock); | |
49 | ||
50 | /* | |
51 | * pnfs_modules_tbl holds all pnfs modules | |
52 | */ | |
53 | static LIST_HEAD(pnfs_modules_tbl); | |
54 | ||
13c13a6a | 55 | static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo); |
aa1e0e3a | 56 | |
02c35fca FI |
57 | /* Return the registered pnfs layout driver module matching given id */ |
58 | static struct pnfs_layoutdriver_type * | |
59 | find_pnfs_driver_locked(u32 id) | |
60 | { | |
61 | struct pnfs_layoutdriver_type *local; | |
62 | ||
63 | list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid) | |
64 | if (local->id == id) | |
65 | goto out; | |
66 | local = NULL; | |
67 | out: | |
68 | dprintk("%s: Searching for id %u, found %p\n", __func__, id, local); | |
69 | return local; | |
70 | } | |
71 | ||
85e174ba RL |
72 | static struct pnfs_layoutdriver_type * |
73 | find_pnfs_driver(u32 id) | |
74 | { | |
02c35fca FI |
75 | struct pnfs_layoutdriver_type *local; |
76 | ||
77 | spin_lock(&pnfs_spinlock); | |
78 | local = find_pnfs_driver_locked(id); | |
0a9c63fa TM |
79 | if (local != NULL && !try_module_get(local->owner)) { |
80 | dprintk("%s: Could not grab reference on module\n", __func__); | |
81 | local = NULL; | |
82 | } | |
02c35fca FI |
83 | spin_unlock(&pnfs_spinlock); |
84 | return local; | |
85e174ba RL |
85 | } |
86 | ||
87 | void | |
88 | unset_pnfs_layoutdriver(struct nfs_server *nfss) | |
89 | { | |
738fd0f3 BH |
90 | if (nfss->pnfs_curr_ld) { |
91 | if (nfss->pnfs_curr_ld->clear_layoutdriver) | |
92 | nfss->pnfs_curr_ld->clear_layoutdriver(nfss); | |
2a4c8994 TM |
93 | /* Decrement the MDS count. Purge the deviceid cache if zero */ |
94 | if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count)) | |
95 | nfs4_deviceid_purge_client(nfss->nfs_client); | |
02c35fca | 96 | module_put(nfss->pnfs_curr_ld->owner); |
738fd0f3 | 97 | } |
85e174ba RL |
98 | nfss->pnfs_curr_ld = NULL; |
99 | } | |
100 | ||
101 | /* | |
102 | * Try to set the server's pnfs module to the pnfs layout type specified by id. | |
103 | * Currently only one pNFS layout driver per filesystem is supported. | |
104 | * | |
105 | * @id layout type. Zero (illegal layout type) indicates pNFS not in use. | |
106 | */ | |
107 | void | |
738fd0f3 BH |
108 | set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh, |
109 | u32 id) | |
85e174ba RL |
110 | { |
111 | struct pnfs_layoutdriver_type *ld_type = NULL; | |
112 | ||
113 | if (id == 0) | |
114 | goto out_no_driver; | |
115 | if (!(server->nfs_client->cl_exchange_flags & | |
116 | (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) { | |
a030889a WAA |
117 | printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n", |
118 | __func__, id, server->nfs_client->cl_exchange_flags); | |
85e174ba RL |
119 | goto out_no_driver; |
120 | } | |
121 | ld_type = find_pnfs_driver(id); | |
122 | if (!ld_type) { | |
123 | request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id); | |
124 | ld_type = find_pnfs_driver(id); | |
125 | if (!ld_type) { | |
126 | dprintk("%s: No pNFS module found for %u.\n", | |
127 | __func__, id); | |
128 | goto out_no_driver; | |
129 | } | |
130 | } | |
131 | server->pnfs_curr_ld = ld_type; | |
738fd0f3 BH |
132 | if (ld_type->set_layoutdriver |
133 | && ld_type->set_layoutdriver(server, mntfh)) { | |
a030889a WAA |
134 | printk(KERN_ERR "NFS: %s: Error initializing pNFS layout " |
135 | "driver %u.\n", __func__, id); | |
738fd0f3 BH |
136 | module_put(ld_type->owner); |
137 | goto out_no_driver; | |
138 | } | |
2a4c8994 TM |
139 | /* Bump the MDS count */ |
140 | atomic_inc(&server->nfs_client->cl_mds_count); | |
ea8eecdd | 141 | |
85e174ba RL |
142 | dprintk("%s: pNFS module for %u set\n", __func__, id); |
143 | return; | |
144 | ||
145 | out_no_driver: | |
146 | dprintk("%s: Using NFSv4 I/O\n", __func__); | |
147 | server->pnfs_curr_ld = NULL; | |
148 | } | |
02c35fca FI |
149 | |
150 | int | |
151 | pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type) | |
152 | { | |
153 | int status = -EINVAL; | |
154 | struct pnfs_layoutdriver_type *tmp; | |
155 | ||
156 | if (ld_type->id == 0) { | |
a030889a | 157 | printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__); |
02c35fca FI |
158 | return status; |
159 | } | |
b1f69b75 | 160 | if (!ld_type->alloc_lseg || !ld_type->free_lseg) { |
a030889a | 161 | printk(KERN_ERR "NFS: %s Layout driver must provide " |
b1f69b75 AA |
162 | "alloc_lseg and free_lseg.\n", __func__); |
163 | return status; | |
164 | } | |
02c35fca FI |
165 | |
166 | spin_lock(&pnfs_spinlock); | |
167 | tmp = find_pnfs_driver_locked(ld_type->id); | |
168 | if (!tmp) { | |
169 | list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl); | |
170 | status = 0; | |
171 | dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id, | |
172 | ld_type->name); | |
173 | } else { | |
a030889a | 174 | printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n", |
02c35fca FI |
175 | __func__, ld_type->id); |
176 | } | |
177 | spin_unlock(&pnfs_spinlock); | |
178 | ||
179 | return status; | |
180 | } | |
181 | EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver); | |
182 | ||
183 | void | |
184 | pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type) | |
185 | { | |
186 | dprintk("%s Deregistering id:%u\n", __func__, ld_type->id); | |
187 | spin_lock(&pnfs_spinlock); | |
188 | list_del(&ld_type->pnfs_tblid); | |
189 | spin_unlock(&pnfs_spinlock); | |
190 | } | |
191 | EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver); | |
e5e94017 | 192 | |
b1f69b75 AA |
193 | /* |
194 | * pNFS client layout cache | |
195 | */ | |
196 | ||
cc6e5340 | 197 | /* Need to hold i_lock if caller does not already hold reference */ |
43f1b3da | 198 | void |
70c3bd2b | 199 | pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo) |
e5e94017 | 200 | { |
cc6e5340 | 201 | atomic_inc(&lo->plh_refcount); |
e5e94017 BH |
202 | } |
203 | ||
636fb9c8 BH |
204 | static struct pnfs_layout_hdr * |
205 | pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags) | |
206 | { | |
207 | struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld; | |
57934278 | 208 | return ld->alloc_layout_hdr(ino, gfp_flags); |
636fb9c8 BH |
209 | } |
210 | ||
211 | static void | |
212 | pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo) | |
213 | { | |
9c626381 TM |
214 | struct nfs_server *server = NFS_SERVER(lo->plh_inode); |
215 | struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld; | |
216 | ||
217 | if (!list_empty(&lo->plh_layouts)) { | |
218 | struct nfs_client *clp = server->nfs_client; | |
219 | ||
220 | spin_lock(&clp->cl_lock); | |
221 | list_del_init(&lo->plh_layouts); | |
222 | spin_unlock(&clp->cl_lock); | |
223 | } | |
9fa40758 | 224 | put_rpccred(lo->plh_lc_cred); |
57934278 | 225 | return ld->free_layout_hdr(lo); |
636fb9c8 BH |
226 | } |
227 | ||
e5e94017 | 228 | static void |
6622c3ea | 229 | pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo) |
e5e94017 | 230 | { |
bb346f63 | 231 | struct nfs_inode *nfsi = NFS_I(lo->plh_inode); |
cc6e5340 | 232 | dprintk("%s: freeing layout cache %p\n", __func__, lo); |
bb346f63 TM |
233 | nfsi->layout = NULL; |
234 | /* Reset MDS Threshold I/O counters */ | |
235 | nfsi->write_io = 0; | |
236 | nfsi->read_io = 0; | |
e5e94017 BH |
237 | } |
238 | ||
b1f69b75 | 239 | void |
70c3bd2b | 240 | pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo) |
974cec8c | 241 | { |
cc6e5340 FI |
242 | struct inode *inode = lo->plh_inode; |
243 | ||
13c13a6a TM |
244 | pnfs_layoutreturn_before_put_layout_hdr(lo); |
245 | ||
cc6e5340 | 246 | if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) { |
566f8737 PT |
247 | if (!list_empty(&lo->plh_segs)) |
248 | WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n"); | |
6622c3ea | 249 | pnfs_detach_layout_hdr(lo); |
cc6e5340 | 250 | spin_unlock(&inode->i_lock); |
6622c3ea | 251 | pnfs_free_layout_hdr(lo); |
cc6e5340 | 252 | } |
974cec8c AA |
253 | } |
254 | ||
2454dfea TM |
255 | /* |
256 | * Mark a pnfs_layout_hdr and all associated layout segments as invalid | |
257 | * | |
258 | * In order to continue using the pnfs_layout_hdr, a full recovery | |
259 | * is required. | |
260 | * Note that caller must hold inode->i_lock. | |
261 | */ | |
9fd4b9fc | 262 | static int |
2454dfea TM |
263 | pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo, |
264 | struct list_head *lseg_list) | |
265 | { | |
266 | struct pnfs_layout_range range = { | |
267 | .iomode = IOMODE_ANY, | |
268 | .offset = 0, | |
269 | .length = NFS4_MAX_UINT64, | |
270 | }; | |
271 | ||
272 | set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags); | |
6d597e17 | 273 | return pnfs_mark_matching_lsegs_invalid(lo, lseg_list, &range, 0); |
2454dfea TM |
274 | } |
275 | ||
b9e028fd TM |
276 | static int |
277 | pnfs_iomode_to_fail_bit(u32 iomode) | |
278 | { | |
279 | return iomode == IOMODE_RW ? | |
280 | NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED; | |
281 | } | |
282 | ||
283 | static void | |
3e621214 | 284 | pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit) |
b9e028fd | 285 | { |
25c75333 | 286 | lo->plh_retry_timestamp = jiffies; |
39e88fcf | 287 | if (!test_and_set_bit(fail_bit, &lo->plh_flags)) |
3e621214 TM |
288 | atomic_inc(&lo->plh_refcount); |
289 | } | |
290 | ||
291 | static void | |
292 | pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit) | |
293 | { | |
294 | if (test_and_clear_bit(fail_bit, &lo->plh_flags)) | |
295 | atomic_dec(&lo->plh_refcount); | |
296 | } | |
297 | ||
298 | static void | |
299 | pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode) | |
300 | { | |
301 | struct inode *inode = lo->plh_inode; | |
115ce575 TM |
302 | struct pnfs_layout_range range = { |
303 | .iomode = iomode, | |
304 | .offset = 0, | |
305 | .length = NFS4_MAX_UINT64, | |
306 | }; | |
307 | LIST_HEAD(head); | |
3e621214 TM |
308 | |
309 | spin_lock(&inode->i_lock); | |
310 | pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode)); | |
6d597e17 | 311 | pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0); |
3e621214 | 312 | spin_unlock(&inode->i_lock); |
115ce575 | 313 | pnfs_free_lseg_list(&head); |
b9e028fd TM |
314 | dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__, |
315 | iomode == IOMODE_RW ? "RW" : "READ"); | |
316 | } | |
317 | ||
318 | static bool | |
319 | pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode) | |
320 | { | |
25c75333 | 321 | unsigned long start, end; |
3e621214 TM |
322 | int fail_bit = pnfs_iomode_to_fail_bit(iomode); |
323 | ||
324 | if (test_bit(fail_bit, &lo->plh_flags) == 0) | |
25c75333 TM |
325 | return false; |
326 | end = jiffies; | |
327 | start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT; | |
328 | if (!time_in_range(lo->plh_retry_timestamp, start, end)) { | |
329 | /* It is time to retry the failed layoutgets */ | |
3e621214 | 330 | pnfs_layout_clear_fail_bit(lo, fail_bit); |
25c75333 TM |
331 | return false; |
332 | } | |
333 | return true; | |
b9e028fd TM |
334 | } |
335 | ||
974cec8c AA |
336 | static void |
337 | init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg) | |
338 | { | |
566052c5 | 339 | INIT_LIST_HEAD(&lseg->pls_list); |
a9bae566 | 340 | INIT_LIST_HEAD(&lseg->pls_lc_list); |
4541d16c FI |
341 | atomic_set(&lseg->pls_refcount, 1); |
342 | smp_mb(); | |
343 | set_bit(NFS_LSEG_VALID, &lseg->pls_flags); | |
566052c5 | 344 | lseg->pls_layout = lo; |
974cec8c AA |
345 | } |
346 | ||
905ca191 | 347 | static void pnfs_free_lseg(struct pnfs_layout_segment *lseg) |
974cec8c | 348 | { |
b7edfaa1 | 349 | struct inode *ino = lseg->pls_layout->plh_inode; |
974cec8c | 350 | |
b1f69b75 | 351 | NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg); |
974cec8c AA |
352 | } |
353 | ||
d684d2ae | 354 | static void |
57036a37 TM |
355 | pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo, |
356 | struct pnfs_layout_segment *lseg) | |
d684d2ae | 357 | { |
57036a37 | 358 | struct inode *inode = lo->plh_inode; |
d684d2ae | 359 | |
d20581aa | 360 | WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags)); |
d684d2ae | 361 | list_del_init(&lseg->pls_list); |
8f0d27dc TM |
362 | /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */ |
363 | atomic_dec(&lo->plh_refcount); | |
173f77e9 TM |
364 | if (list_empty(&lo->plh_segs)) |
365 | clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags); | |
d684d2ae FI |
366 | rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq); |
367 | } | |
368 | ||
bae724ef | 369 | void |
9369a431 | 370 | pnfs_put_lseg(struct pnfs_layout_segment *lseg) |
974cec8c | 371 | { |
57036a37 | 372 | struct pnfs_layout_hdr *lo; |
d684d2ae FI |
373 | struct inode *inode; |
374 | ||
375 | if (!lseg) | |
376 | return; | |
377 | ||
4541d16c FI |
378 | dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg, |
379 | atomic_read(&lseg->pls_refcount), | |
380 | test_bit(NFS_LSEG_VALID, &lseg->pls_flags)); | |
4ef2e4f8 | 381 | |
57036a37 TM |
382 | lo = lseg->pls_layout; |
383 | inode = lo->plh_inode; | |
4ef2e4f8 | 384 | |
d684d2ae | 385 | if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) { |
faa4a54f TM |
386 | if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) { |
387 | spin_unlock(&inode->i_lock); | |
388 | return; | |
389 | } | |
8f0d27dc | 390 | pnfs_get_layout_hdr(lo); |
4ef2e4f8 TM |
391 | pnfs_layout_remove_lseg(lo, lseg); |
392 | spin_unlock(&inode->i_lock); | |
393 | pnfs_free_lseg(lseg); | |
394 | pnfs_put_layout_hdr(lo); | |
4541d16c | 395 | } |
4541d16c | 396 | } |
9369a431 | 397 | EXPORT_SYMBOL_GPL(pnfs_put_lseg); |
974cec8c | 398 | |
6543f803 | 399 | static void pnfs_free_lseg_async_work(struct work_struct *work) |
e6cf82d1 WAA |
400 | { |
401 | struct pnfs_layout_segment *lseg; | |
6543f803 | 402 | struct pnfs_layout_hdr *lo; |
e6cf82d1 WAA |
403 | |
404 | lseg = container_of(work, struct pnfs_layout_segment, pls_work); | |
6543f803 | 405 | lo = lseg->pls_layout; |
e6cf82d1 | 406 | |
6543f803 TM |
407 | pnfs_free_lseg(lseg); |
408 | pnfs_put_layout_hdr(lo); | |
e6cf82d1 WAA |
409 | } |
410 | ||
6543f803 | 411 | static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg) |
e6cf82d1 | 412 | { |
6543f803 | 413 | INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work); |
e6cf82d1 WAA |
414 | schedule_work(&lseg->pls_work); |
415 | } | |
6543f803 TM |
416 | |
417 | void | |
418 | pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg) | |
419 | { | |
420 | if (!lseg) | |
421 | return; | |
422 | ||
423 | assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock); | |
424 | ||
425 | dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg, | |
426 | atomic_read(&lseg->pls_refcount), | |
427 | test_bit(NFS_LSEG_VALID, &lseg->pls_flags)); | |
428 | if (atomic_dec_and_test(&lseg->pls_refcount)) { | |
429 | struct pnfs_layout_hdr *lo = lseg->pls_layout; | |
faa4a54f TM |
430 | if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) |
431 | return; | |
6543f803 TM |
432 | pnfs_get_layout_hdr(lo); |
433 | pnfs_layout_remove_lseg(lo, lseg); | |
434 | pnfs_free_lseg_async(lseg); | |
435 | } | |
436 | } | |
437 | EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked); | |
e6cf82d1 | 438 | |
3cb2df17 | 439 | static u64 |
fb3296eb BH |
440 | end_offset(u64 start, u64 len) |
441 | { | |
442 | u64 end; | |
443 | ||
444 | end = start + len; | |
445 | return end >= start ? end : NFS4_MAX_UINT64; | |
446 | } | |
447 | ||
fb3296eb BH |
448 | /* |
449 | * is l2 fully contained in l1? | |
450 | * start1 end1 | |
451 | * [----------------------------------) | |
452 | * start2 end2 | |
453 | * [----------------) | |
454 | */ | |
3cb2df17 | 455 | static bool |
7dc0ac70 | 456 | pnfs_lseg_range_contained(const struct pnfs_layout_range *l1, |
3cb2df17 | 457 | const struct pnfs_layout_range *l2) |
fb3296eb BH |
458 | { |
459 | u64 start1 = l1->offset; | |
460 | u64 end1 = end_offset(start1, l1->length); | |
461 | u64 start2 = l2->offset; | |
462 | u64 end2 = end_offset(start2, l2->length); | |
463 | ||
464 | return (start1 <= start2) && (end1 >= end2); | |
465 | } | |
466 | ||
467 | /* | |
468 | * is l1 and l2 intersecting? | |
469 | * start1 end1 | |
470 | * [----------------------------------) | |
471 | * start2 end2 | |
472 | * [----------------) | |
473 | */ | |
3cb2df17 | 474 | static bool |
7dc0ac70 | 475 | pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1, |
3cb2df17 | 476 | const struct pnfs_layout_range *l2) |
fb3296eb BH |
477 | { |
478 | u64 start1 = l1->offset; | |
479 | u64 end1 = end_offset(start1, l1->length); | |
480 | u64 start2 = l2->offset; | |
481 | u64 end2 = end_offset(start2, l2->length); | |
482 | ||
483 | return (end1 == NFS4_MAX_UINT64 || end1 > start2) && | |
484 | (end2 == NFS4_MAX_UINT64 || end2 > start1); | |
485 | } | |
486 | ||
4541d16c | 487 | static bool |
3cb2df17 TM |
488 | should_free_lseg(const struct pnfs_layout_range *lseg_range, |
489 | const struct pnfs_layout_range *recall_range) | |
4541d16c | 490 | { |
778b5502 BH |
491 | return (recall_range->iomode == IOMODE_ANY || |
492 | lseg_range->iomode == recall_range->iomode) && | |
7dc0ac70 | 493 | pnfs_lseg_range_intersecting(lseg_range, recall_range); |
974cec8c AA |
494 | } |
495 | ||
24956804 TM |
496 | static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg, |
497 | struct list_head *tmp_list) | |
498 | { | |
499 | if (!atomic_dec_and_test(&lseg->pls_refcount)) | |
500 | return false; | |
501 | pnfs_layout_remove_lseg(lseg->pls_layout, lseg); | |
502 | list_add(&lseg->pls_list, tmp_list); | |
503 | return true; | |
504 | } | |
505 | ||
4541d16c FI |
506 | /* Returns 1 if lseg is removed from list, 0 otherwise */ |
507 | static int mark_lseg_invalid(struct pnfs_layout_segment *lseg, | |
508 | struct list_head *tmp_list) | |
509 | { | |
510 | int rv = 0; | |
511 | ||
512 | if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) { | |
513 | /* Remove the reference keeping the lseg in the | |
514 | * list. It will now be removed when all | |
515 | * outstanding io is finished. | |
516 | */ | |
d684d2ae FI |
517 | dprintk("%s: lseg %p ref %d\n", __func__, lseg, |
518 | atomic_read(&lseg->pls_refcount)); | |
24956804 | 519 | if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list)) |
d684d2ae | 520 | rv = 1; |
4541d16c FI |
521 | } |
522 | return rv; | |
523 | } | |
524 | ||
6d597e17 JL |
525 | /* |
526 | * Compare 2 layout stateid sequence ids, to see which is newer, | |
527 | * taking into account wraparound issues. | |
528 | */ | |
529 | static bool pnfs_seqid_is_newer(u32 s1, u32 s2) | |
530 | { | |
531 | return (s32)(s1 - s2) > 0; | |
532 | } | |
533 | ||
534 | /** | |
535 | * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later | |
536 | * @lo: layout header containing the lsegs | |
537 | * @tmp_list: list head where doomed lsegs should go | |
538 | * @recall_range: optional recall range argument to match (may be NULL) | |
539 | * @seq: only invalidate lsegs obtained prior to this sequence (may be 0) | |
540 | * | |
541 | * Walk the list of lsegs in the layout header, and tear down any that should | |
542 | * be destroyed. If "recall_range" is specified then the segment must match | |
543 | * that range. If "seq" is non-zero, then only match segments that were handed | |
544 | * out at or before that sequence. | |
545 | * | |
546 | * Returns number of matching invalid lsegs remaining in list after scanning | |
547 | * it and purging them. | |
4541d16c | 548 | */ |
43f1b3da | 549 | int |
49a85061 | 550 | pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo, |
4541d16c | 551 | struct list_head *tmp_list, |
6d597e17 JL |
552 | const struct pnfs_layout_range *recall_range, |
553 | u32 seq) | |
974cec8c AA |
554 | { |
555 | struct pnfs_layout_segment *lseg, *next; | |
71b39854 | 556 | int remaining = 0; |
974cec8c AA |
557 | |
558 | dprintk("%s:Begin lo %p\n", __func__, lo); | |
559 | ||
8006bfba | 560 | if (list_empty(&lo->plh_segs)) |
38511722 | 561 | return 0; |
4541d16c | 562 | list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) |
778b5502 BH |
563 | if (!recall_range || |
564 | should_free_lseg(&lseg->pls_range, recall_range)) { | |
6d597e17 JL |
565 | if (seq && pnfs_seqid_is_newer(lseg->pls_seq, seq)) |
566 | continue; | |
567 | dprintk("%s: freeing lseg %p iomode %d seq %u" | |
4541d16c | 568 | "offset %llu length %llu\n", __func__, |
6d597e17 JL |
569 | lseg, lseg->pls_range.iomode, lseg->pls_seq, |
570 | lseg->pls_range.offset, lseg->pls_range.length); | |
71b39854 TM |
571 | if (!mark_lseg_invalid(lseg, tmp_list)) |
572 | remaining++; | |
4541d16c | 573 | } |
71b39854 TM |
574 | dprintk("%s:Return %i\n", __func__, remaining); |
575 | return remaining; | |
974cec8c AA |
576 | } |
577 | ||
f49f9baa | 578 | /* note free_me must contain lsegs from a single layout_hdr */ |
43f1b3da | 579 | void |
4541d16c | 580 | pnfs_free_lseg_list(struct list_head *free_me) |
974cec8c | 581 | { |
4541d16c | 582 | struct pnfs_layout_segment *lseg, *tmp; |
f49f9baa FI |
583 | |
584 | if (list_empty(free_me)) | |
585 | return; | |
586 | ||
4541d16c | 587 | list_for_each_entry_safe(lseg, tmp, free_me, pls_list) { |
566052c5 | 588 | list_del(&lseg->pls_list); |
905ca191 | 589 | pnfs_free_lseg(lseg); |
974cec8c AA |
590 | } |
591 | } | |
592 | ||
e5e94017 BH |
593 | void |
594 | pnfs_destroy_layout(struct nfs_inode *nfsi) | |
595 | { | |
596 | struct pnfs_layout_hdr *lo; | |
974cec8c | 597 | LIST_HEAD(tmp_list); |
e5e94017 BH |
598 | |
599 | spin_lock(&nfsi->vfs_inode.i_lock); | |
600 | lo = nfsi->layout; | |
601 | if (lo) { | |
3e621214 | 602 | pnfs_get_layout_hdr(lo); |
2454dfea | 603 | pnfs_mark_layout_stateid_invalid(lo, &tmp_list); |
3e621214 TM |
604 | pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED); |
605 | pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED); | |
606 | spin_unlock(&nfsi->vfs_inode.i_lock); | |
607 | pnfs_free_lseg_list(&tmp_list); | |
608 | pnfs_put_layout_hdr(lo); | |
609 | } else | |
610 | spin_unlock(&nfsi->vfs_inode.i_lock); | |
974cec8c | 611 | } |
041245c8 | 612 | EXPORT_SYMBOL_GPL(pnfs_destroy_layout); |
974cec8c | 613 | |
fd9a8d71 TM |
614 | static bool |
615 | pnfs_layout_add_bulk_destroy_list(struct inode *inode, | |
616 | struct list_head *layout_list) | |
974cec8c AA |
617 | { |
618 | struct pnfs_layout_hdr *lo; | |
fd9a8d71 | 619 | bool ret = false; |
974cec8c | 620 | |
fd9a8d71 TM |
621 | spin_lock(&inode->i_lock); |
622 | lo = NFS_I(inode)->layout; | |
623 | if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) { | |
624 | pnfs_get_layout_hdr(lo); | |
625 | list_add(&lo->plh_bulk_destroy, layout_list); | |
626 | ret = true; | |
627 | } | |
628 | spin_unlock(&inode->i_lock); | |
629 | return ret; | |
630 | } | |
631 | ||
632 | /* Caller must hold rcu_read_lock and clp->cl_lock */ | |
633 | static int | |
634 | pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp, | |
635 | struct nfs_server *server, | |
636 | struct list_head *layout_list) | |
637 | { | |
638 | struct pnfs_layout_hdr *lo, *next; | |
639 | struct inode *inode; | |
640 | ||
641 | list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) { | |
642 | inode = igrab(lo->plh_inode); | |
643 | if (inode == NULL) | |
644 | continue; | |
645 | list_del_init(&lo->plh_layouts); | |
646 | if (pnfs_layout_add_bulk_destroy_list(inode, layout_list)) | |
647 | continue; | |
648 | rcu_read_unlock(); | |
649 | spin_unlock(&clp->cl_lock); | |
650 | iput(inode); | |
651 | spin_lock(&clp->cl_lock); | |
652 | rcu_read_lock(); | |
653 | return -EAGAIN; | |
654 | } | |
655 | return 0; | |
656 | } | |
657 | ||
658 | static int | |
659 | pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list, | |
660 | bool is_bulk_recall) | |
661 | { | |
662 | struct pnfs_layout_hdr *lo; | |
663 | struct inode *inode; | |
fd9a8d71 TM |
664 | LIST_HEAD(lseg_list); |
665 | int ret = 0; | |
666 | ||
667 | while (!list_empty(layout_list)) { | |
668 | lo = list_entry(layout_list->next, struct pnfs_layout_hdr, | |
669 | plh_bulk_destroy); | |
670 | dprintk("%s freeing layout for inode %lu\n", __func__, | |
671 | lo->plh_inode->i_ino); | |
672 | inode = lo->plh_inode; | |
7c5d1875 CH |
673 | |
674 | pnfs_layoutcommit_inode(inode, false); | |
675 | ||
fd9a8d71 TM |
676 | spin_lock(&inode->i_lock); |
677 | list_del_init(&lo->plh_bulk_destroy); | |
9fd4b9fc TM |
678 | if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) { |
679 | if (is_bulk_recall) | |
680 | set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags); | |
fd9a8d71 | 681 | ret = -EAGAIN; |
9fd4b9fc | 682 | } |
fd9a8d71 TM |
683 | spin_unlock(&inode->i_lock); |
684 | pnfs_free_lseg_list(&lseg_list); | |
b20135d0 TM |
685 | /* Free all lsegs that are attached to commit buckets */ |
686 | nfs_commit_inode(inode, 0); | |
fd9a8d71 TM |
687 | pnfs_put_layout_hdr(lo); |
688 | iput(inode); | |
689 | } | |
690 | return ret; | |
691 | } | |
692 | ||
693 | int | |
694 | pnfs_destroy_layouts_byfsid(struct nfs_client *clp, | |
695 | struct nfs_fsid *fsid, | |
696 | bool is_recall) | |
697 | { | |
698 | struct nfs_server *server; | |
699 | LIST_HEAD(layout_list); | |
c47abcf8 | 700 | |
974cec8c | 701 | spin_lock(&clp->cl_lock); |
6382a441 | 702 | rcu_read_lock(); |
fd9a8d71 | 703 | restart: |
6382a441 | 704 | list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) { |
fd9a8d71 TM |
705 | if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0) |
706 | continue; | |
707 | if (pnfs_layout_bulk_destroy_byserver_locked(clp, | |
708 | server, | |
709 | &layout_list) != 0) | |
710 | goto restart; | |
6382a441 WAA |
711 | } |
712 | rcu_read_unlock(); | |
974cec8c AA |
713 | spin_unlock(&clp->cl_lock); |
714 | ||
fd9a8d71 TM |
715 | if (list_empty(&layout_list)) |
716 | return 0; | |
717 | return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall); | |
718 | } | |
719 | ||
720 | int | |
721 | pnfs_destroy_layouts_byclid(struct nfs_client *clp, | |
722 | bool is_recall) | |
723 | { | |
724 | struct nfs_server *server; | |
725 | LIST_HEAD(layout_list); | |
726 | ||
727 | spin_lock(&clp->cl_lock); | |
728 | rcu_read_lock(); | |
729 | restart: | |
730 | list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) { | |
731 | if (pnfs_layout_bulk_destroy_byserver_locked(clp, | |
732 | server, | |
733 | &layout_list) != 0) | |
734 | goto restart; | |
974cec8c | 735 | } |
fd9a8d71 TM |
736 | rcu_read_unlock(); |
737 | spin_unlock(&clp->cl_lock); | |
738 | ||
739 | if (list_empty(&layout_list)) | |
740 | return 0; | |
741 | return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall); | |
742 | } | |
743 | ||
744 | /* | |
745 | * Called by the state manger to remove all layouts established under an | |
746 | * expired lease. | |
747 | */ | |
748 | void | |
749 | pnfs_destroy_all_layouts(struct nfs_client *clp) | |
750 | { | |
751 | nfs4_deviceid_mark_client_invalid(clp); | |
752 | nfs4_deviceid_purge_client(clp); | |
753 | ||
754 | pnfs_destroy_layouts_byclid(clp, false); | |
e5e94017 BH |
755 | } |
756 | ||
fd6002e9 | 757 | /* update lo->plh_stateid with new if is more recent */ |
43f1b3da FI |
758 | void |
759 | pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new, | |
760 | bool update_barrier) | |
b1f69b75 | 761 | { |
22aaf714 TM |
762 | u32 oldseq, newseq, new_barrier; |
763 | int empty = list_empty(&lo->plh_segs); | |
b1f69b75 | 764 | |
2d2f24ad TM |
765 | oldseq = be32_to_cpu(lo->plh_stateid.seqid); |
766 | newseq = be32_to_cpu(new->seqid); | |
22aaf714 | 767 | if (empty || pnfs_seqid_is_newer(newseq, oldseq)) { |
f597c537 | 768 | nfs4_stateid_copy(&lo->plh_stateid, new); |
43f1b3da | 769 | if (update_barrier) { |
22aaf714 | 770 | new_barrier = be32_to_cpu(new->seqid); |
43f1b3da FI |
771 | } else { |
772 | /* Because of wraparound, we want to keep the barrier | |
22aaf714 | 773 | * "close" to the current seqids. |
43f1b3da | 774 | */ |
22aaf714 | 775 | new_barrier = newseq - atomic_read(&lo->plh_outstanding); |
43f1b3da | 776 | } |
22aaf714 TM |
777 | if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier)) |
778 | lo->plh_barrier = new_barrier; | |
43f1b3da | 779 | } |
b1f69b75 AA |
780 | } |
781 | ||
cf7d63f1 | 782 | static bool |
19c54aba TM |
783 | pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo, |
784 | const nfs4_stateid *stateid) | |
43f1b3da | 785 | { |
19c54aba | 786 | u32 seqid = be32_to_cpu(stateid->seqid); |
25a1a621 | 787 | |
19c54aba TM |
788 | return !pnfs_seqid_is_newer(seqid, lo->plh_barrier); |
789 | } | |
790 | ||
791 | /* lget is set to 1 if called from inside send_layoutget call chain */ | |
792 | static bool | |
e1c06f80 | 793 | pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo) |
19c54aba | 794 | { |
f7e8917a | 795 | return lo->plh_block_lgets || |
e1c06f80 | 796 | test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags); |
cf7d63f1 FI |
797 | } |
798 | ||
fd6002e9 FI |
799 | int |
800 | pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo, | |
506c0d68 | 801 | const struct pnfs_layout_range *range, |
fd6002e9 | 802 | struct nfs4_state *open_state) |
b1f69b75 | 803 | { |
fd6002e9 | 804 | int status = 0; |
974cec8c | 805 | |
b1f69b75 | 806 | dprintk("--> %s\n", __func__); |
fd6002e9 | 807 | spin_lock(&lo->plh_inode->i_lock); |
e1c06f80 | 808 | if (pnfs_layoutgets_blocked(lo)) { |
cf7d63f1 | 809 | status = -EAGAIN; |
5d422301 TM |
810 | } else if (!nfs4_valid_open_stateid(open_state)) { |
811 | status = -EBADF; | |
47abadef CH |
812 | } else if (list_empty(&lo->plh_segs) || |
813 | test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) { | |
fd6002e9 FI |
814 | int seq; |
815 | ||
816 | do { | |
817 | seq = read_seqbegin(&open_state->seqlock); | |
f597c537 | 818 | nfs4_stateid_copy(dst, &open_state->stateid); |
fd6002e9 FI |
819 | } while (read_seqretry(&open_state->seqlock, seq)); |
820 | } else | |
f597c537 | 821 | nfs4_stateid_copy(dst, &lo->plh_stateid); |
fd6002e9 | 822 | spin_unlock(&lo->plh_inode->i_lock); |
b1f69b75 | 823 | dprintk("<-- %s\n", __func__); |
fd6002e9 | 824 | return status; |
b1f69b75 AA |
825 | } |
826 | ||
827 | /* | |
828 | * Get layout from server. | |
829 | * for now, assume that whole file layouts are requested. | |
830 | * arg->offset: 0 | |
831 | * arg->length: all ones | |
832 | */ | |
e5e94017 BH |
833 | static struct pnfs_layout_segment * |
834 | send_layoutget(struct pnfs_layout_hdr *lo, | |
835 | struct nfs_open_context *ctx, | |
e144e539 | 836 | const struct pnfs_layout_range *range, |
a75b9df9 | 837 | gfp_t gfp_flags) |
e5e94017 | 838 | { |
b7edfaa1 | 839 | struct inode *ino = lo->plh_inode; |
b1f69b75 AA |
840 | struct nfs_server *server = NFS_SERVER(ino); |
841 | struct nfs4_layoutget *lgp; | |
a0b0a6e3 | 842 | struct pnfs_layout_segment *lseg; |
2d89a1d3 | 843 | loff_t i_size; |
b1f69b75 AA |
844 | |
845 | dprintk("--> %s\n", __func__); | |
e5e94017 | 846 | |
4f2e9dce JL |
847 | /* |
848 | * Synchronously retrieve layout information from server and | |
849 | * store in lseg. If we race with a concurrent seqid morphing | |
850 | * op, then re-send the LAYOUTGET. | |
851 | */ | |
852 | do { | |
853 | lgp = kzalloc(sizeof(*lgp), gfp_flags); | |
854 | if (lgp == NULL) | |
855 | return NULL; | |
856 | ||
857 | i_size = i_size_read(ino); | |
858 | ||
09cbfeaf | 859 | lgp->args.minlength = PAGE_SIZE; |
4f2e9dce JL |
860 | if (lgp->args.minlength > range->length) |
861 | lgp->args.minlength = range->length; | |
862 | if (range->iomode == IOMODE_READ) { | |
863 | if (range->offset >= i_size) | |
864 | lgp->args.minlength = 0; | |
865 | else if (i_size - range->offset < lgp->args.minlength) | |
866 | lgp->args.minlength = i_size - range->offset; | |
867 | } | |
868 | lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE; | |
e144e539 | 869 | pnfs_copy_range(&lgp->args.range, range); |
4f2e9dce JL |
870 | lgp->args.type = server->pnfs_curr_ld->id; |
871 | lgp->args.inode = ino; | |
872 | lgp->args.ctx = get_nfs_open_context(ctx); | |
873 | lgp->gfp_flags = gfp_flags; | |
874 | lgp->cred = lo->plh_lc_cred; | |
35124a09 | 875 | |
4f2e9dce JL |
876 | lseg = nfs4_proc_layoutget(lgp, gfp_flags); |
877 | } while (lseg == ERR_PTR(-EAGAIN)); | |
2d89a1d3 | 878 | |
d03ab29d JL |
879 | if (IS_ERR(lseg)) { |
880 | if (!nfs_error_is_fatal(PTR_ERR(lseg))) { | |
881 | pnfs_layout_clear_fail_bit(lo, | |
882 | pnfs_iomode_to_fail_bit(range->iomode)); | |
883 | lseg = NULL; | |
884 | } | |
885 | } else { | |
d67ae825 TH |
886 | pnfs_layout_clear_fail_bit(lo, |
887 | pnfs_iomode_to_fail_bit(range->iomode)); | |
d03ab29d | 888 | } |
35124a09 | 889 | |
974cec8c AA |
890 | return lseg; |
891 | } | |
892 | ||
24956804 TM |
893 | static void pnfs_clear_layoutcommit(struct inode *inode, |
894 | struct list_head *head) | |
895 | { | |
896 | struct nfs_inode *nfsi = NFS_I(inode); | |
897 | struct pnfs_layout_segment *lseg, *tmp; | |
898 | ||
899 | if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) | |
900 | return; | |
901 | list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) { | |
902 | if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) | |
903 | continue; | |
904 | pnfs_lseg_dec_and_remove_zero(lseg, head); | |
905 | } | |
906 | } | |
907 | ||
d67ae825 TH |
908 | void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo) |
909 | { | |
910 | clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags); | |
911 | smp_mb__after_atomic(); | |
912 | wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN); | |
7f27392c | 913 | rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq); |
d67ae825 TH |
914 | } |
915 | ||
13c13a6a TM |
916 | static bool |
917 | pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo) | |
918 | { | |
919 | if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) | |
920 | return false; | |
921 | lo->plh_return_iomode = 0; | |
3982a6a2 | 922 | lo->plh_return_seq = 0; |
13c13a6a | 923 | pnfs_get_layout_hdr(lo); |
2370abda | 924 | clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags); |
13c13a6a TM |
925 | return true; |
926 | } | |
927 | ||
f40eb5d0 | 928 | static int |
ed429d6b | 929 | pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid, |
6c16605d | 930 | enum pnfs_iomode iomode, bool sync) |
f40eb5d0 PT |
931 | { |
932 | struct inode *ino = lo->plh_inode; | |
933 | struct nfs4_layoutreturn *lrp; | |
934 | int status = 0; | |
935 | ||
e4af440a | 936 | lrp = kzalloc(sizeof(*lrp), GFP_NOFS); |
f40eb5d0 PT |
937 | if (unlikely(lrp == NULL)) { |
938 | status = -ENOMEM; | |
939 | spin_lock(&ino->i_lock); | |
d67ae825 | 940 | pnfs_clear_layoutreturn_waitbit(lo); |
f40eb5d0 PT |
941 | spin_unlock(&ino->i_lock); |
942 | pnfs_put_layout_hdr(lo); | |
943 | goto out; | |
944 | } | |
945 | ||
ed429d6b | 946 | nfs4_stateid_copy(&lrp->args.stateid, stateid); |
f40eb5d0 PT |
947 | lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id; |
948 | lrp->args.inode = ino; | |
15eb67c1 PT |
949 | lrp->args.range.iomode = iomode; |
950 | lrp->args.range.offset = 0; | |
951 | lrp->args.range.length = NFS4_MAX_UINT64; | |
f40eb5d0 PT |
952 | lrp->args.layout = lo; |
953 | lrp->clp = NFS_SERVER(ino)->nfs_client; | |
954 | lrp->cred = lo->plh_lc_cred; | |
955 | ||
6c16605d | 956 | status = nfs4_proc_layoutreturn(lrp, sync); |
f40eb5d0 PT |
957 | out: |
958 | dprintk("<-- %s status: %d\n", __func__, status); | |
959 | return status; | |
960 | } | |
961 | ||
13c13a6a TM |
962 | /* Return true if layoutreturn is needed */ |
963 | static bool | |
964 | pnfs_layout_need_return(struct pnfs_layout_hdr *lo) | |
965 | { | |
966 | struct pnfs_layout_segment *s; | |
967 | ||
2370abda | 968 | if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) |
13c13a6a TM |
969 | return false; |
970 | ||
971 | /* Defer layoutreturn until all lsegs are done */ | |
972 | list_for_each_entry(s, &lo->plh_segs, pls_list) { | |
973 | if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags)) | |
974 | return false; | |
975 | } | |
976 | ||
977 | return true; | |
978 | } | |
979 | ||
980 | static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo) | |
981 | { | |
982 | struct inode *inode= lo->plh_inode; | |
983 | ||
2370abda | 984 | if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) |
13c13a6a TM |
985 | return; |
986 | spin_lock(&inode->i_lock); | |
987 | if (pnfs_layout_need_return(lo)) { | |
988 | nfs4_stateid stateid; | |
989 | enum pnfs_iomode iomode; | |
990 | bool send; | |
991 | ||
992 | nfs4_stateid_copy(&stateid, &lo->plh_stateid); | |
3982a6a2 | 993 | stateid.seqid = cpu_to_be32(lo->plh_return_seq); |
13c13a6a TM |
994 | iomode = lo->plh_return_iomode; |
995 | send = pnfs_prepare_layoutreturn(lo); | |
996 | spin_unlock(&inode->i_lock); | |
997 | if (send) { | |
998 | /* Send an async layoutreturn so we dont deadlock */ | |
999 | pnfs_send_layoutreturn(lo, &stateid, iomode, false); | |
1000 | } | |
1001 | } else | |
1002 | spin_unlock(&inode->i_lock); | |
1003 | } | |
1004 | ||
293b3b06 AA |
1005 | /* |
1006 | * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr | |
1007 | * when the layout segment list is empty. | |
1008 | * | |
1009 | * Note that a pnfs_layout_hdr can exist with an empty layout segment | |
1010 | * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the | |
1011 | * deviceid is marked invalid. | |
1012 | */ | |
cbe82603 BH |
1013 | int |
1014 | _pnfs_return_layout(struct inode *ino) | |
1015 | { | |
1016 | struct pnfs_layout_hdr *lo = NULL; | |
1017 | struct nfs_inode *nfsi = NFS_I(ino); | |
1018 | LIST_HEAD(tmp_list); | |
cbe82603 | 1019 | nfs4_stateid stateid; |
293b3b06 | 1020 | int status = 0, empty; |
7f27392c | 1021 | bool send; |
cbe82603 | 1022 | |
366d5052 | 1023 | dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino); |
cbe82603 BH |
1024 | |
1025 | spin_lock(&ino->i_lock); | |
1026 | lo = nfsi->layout; | |
e5929f3c | 1027 | if (!lo) { |
cbe82603 | 1028 | spin_unlock(&ino->i_lock); |
293b3b06 AA |
1029 | dprintk("NFS: %s no layout to return\n", __func__); |
1030 | goto out; | |
cbe82603 | 1031 | } |
50f563ef | 1032 | nfs4_stateid_copy(&stateid, &nfsi->layout->plh_stateid); |
cbe82603 | 1033 | /* Reference matched in nfs4_layoutreturn_release */ |
70c3bd2b | 1034 | pnfs_get_layout_hdr(lo); |
293b3b06 | 1035 | empty = list_empty(&lo->plh_segs); |
24956804 | 1036 | pnfs_clear_layoutcommit(ino, &tmp_list); |
6d597e17 | 1037 | pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0); |
c88953d8 CH |
1038 | |
1039 | if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) { | |
1040 | struct pnfs_layout_range range = { | |
1041 | .iomode = IOMODE_ANY, | |
1042 | .offset = 0, | |
1043 | .length = NFS4_MAX_UINT64, | |
1044 | }; | |
1045 | NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range); | |
1046 | } | |
1047 | ||
293b3b06 AA |
1048 | /* Don't send a LAYOUTRETURN if list was initially empty */ |
1049 | if (empty) { | |
1050 | spin_unlock(&ino->i_lock); | |
293b3b06 | 1051 | dprintk("NFS: %s no layout segments to return\n", __func__); |
7f27392c | 1052 | goto out_put_layout_hdr; |
293b3b06 | 1053 | } |
47abadef CH |
1054 | |
1055 | set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags); | |
7f27392c | 1056 | send = pnfs_prepare_layoutreturn(lo); |
cbe82603 BH |
1057 | spin_unlock(&ino->i_lock); |
1058 | pnfs_free_lseg_list(&tmp_list); | |
7f27392c | 1059 | if (send) |
ed429d6b | 1060 | status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true); |
7f27392c TM |
1061 | out_put_layout_hdr: |
1062 | pnfs_put_layout_hdr(lo); | |
cbe82603 BH |
1063 | out: |
1064 | dprintk("<-- %s status: %d\n", __func__, status); | |
1065 | return status; | |
1066 | } | |
0a57cdac | 1067 | EXPORT_SYMBOL_GPL(_pnfs_return_layout); |
cbe82603 | 1068 | |
24028672 TM |
1069 | int |
1070 | pnfs_commit_and_return_layout(struct inode *inode) | |
1071 | { | |
1072 | struct pnfs_layout_hdr *lo; | |
1073 | int ret; | |
1074 | ||
1075 | spin_lock(&inode->i_lock); | |
1076 | lo = NFS_I(inode)->layout; | |
1077 | if (lo == NULL) { | |
1078 | spin_unlock(&inode->i_lock); | |
1079 | return 0; | |
1080 | } | |
1081 | pnfs_get_layout_hdr(lo); | |
1082 | /* Block new layoutgets and read/write to ds */ | |
1083 | lo->plh_block_lgets++; | |
1084 | spin_unlock(&inode->i_lock); | |
1085 | filemap_fdatawait(inode->i_mapping); | |
1086 | ret = pnfs_layoutcommit_inode(inode, true); | |
1087 | if (ret == 0) | |
1088 | ret = _pnfs_return_layout(inode); | |
1089 | spin_lock(&inode->i_lock); | |
1090 | lo->plh_block_lgets--; | |
1091 | spin_unlock(&inode->i_lock); | |
1092 | pnfs_put_layout_hdr(lo); | |
1093 | return ret; | |
1094 | } | |
1095 | ||
f7e8917a FI |
1096 | bool pnfs_roc(struct inode *ino) |
1097 | { | |
40dd4b7a TM |
1098 | struct nfs_inode *nfsi = NFS_I(ino); |
1099 | struct nfs_open_context *ctx; | |
1100 | struct nfs4_state *state; | |
f7e8917a FI |
1101 | struct pnfs_layout_hdr *lo; |
1102 | struct pnfs_layout_segment *lseg, *tmp; | |
193e3aa2 | 1103 | nfs4_stateid stateid; |
f7e8917a | 1104 | LIST_HEAD(tmp_list); |
e755d638 | 1105 | bool found = false, layoutreturn = false, roc = false; |
f7e8917a FI |
1106 | |
1107 | spin_lock(&ino->i_lock); | |
40dd4b7a | 1108 | lo = nfsi->layout; |
3976143b | 1109 | if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) |
40dd4b7a TM |
1110 | goto out_noroc; |
1111 | ||
e755d638 | 1112 | /* no roc if we hold a delegation */ |
40dd4b7a TM |
1113 | if (nfs4_check_delegation(ino, FMODE_READ)) |
1114 | goto out_noroc; | |
1115 | ||
1116 | list_for_each_entry(ctx, &nfsi->open_files, list) { | |
1117 | state = ctx->state; | |
1118 | /* Don't return layout if there is open file state */ | |
1119 | if (state != NULL && state->state != 0) | |
1120 | goto out_noroc; | |
1121 | } | |
1122 | ||
50f563ef | 1123 | nfs4_stateid_copy(&stateid, &lo->plh_stateid); |
e755d638 | 1124 | /* always send layoutreturn if being marked so */ |
2370abda | 1125 | if (test_and_clear_bit(NFS_LAYOUT_RETURN_REQUESTED, |
e755d638 PT |
1126 | &lo->plh_flags)) |
1127 | layoutreturn = pnfs_prepare_layoutreturn(lo); | |
1128 | ||
f7e8917a | 1129 | list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list) |
e755d638 PT |
1130 | /* If we are sending layoutreturn, invalidate all valid lsegs */ |
1131 | if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) { | |
f7e8917a FI |
1132 | mark_lseg_invalid(lseg, &tmp_list); |
1133 | found = true; | |
1134 | } | |
500d701f | 1135 | /* ROC in two conditions: |
e755d638 PT |
1136 | * 1. there are ROC lsegs |
1137 | * 2. we don't send layoutreturn | |
e755d638 | 1138 | */ |
500d701f PT |
1139 | if (found && !layoutreturn) { |
1140 | /* lo ref dropped in pnfs_roc_release() */ | |
1141 | pnfs_get_layout_hdr(lo); | |
e755d638 | 1142 | roc = true; |
500d701f | 1143 | } |
f7e8917a | 1144 | |
40dd4b7a | 1145 | out_noroc: |
f7e8917a | 1146 | spin_unlock(&ino->i_lock); |
e755d638 PT |
1147 | pnfs_free_lseg_list(&tmp_list); |
1148 | pnfs_layoutcommit_inode(ino, true); | |
1149 | if (layoutreturn) | |
ed429d6b | 1150 | pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true); |
e755d638 | 1151 | return roc; |
f7e8917a FI |
1152 | } |
1153 | ||
1154 | void pnfs_roc_release(struct inode *ino) | |
1155 | { | |
1156 | struct pnfs_layout_hdr *lo; | |
1157 | ||
1158 | spin_lock(&ino->i_lock); | |
1159 | lo = NFS_I(ino)->layout; | |
5c4a79fb | 1160 | pnfs_clear_layoutreturn_waitbit(lo); |
6622c3ea TM |
1161 | if (atomic_dec_and_test(&lo->plh_refcount)) { |
1162 | pnfs_detach_layout_hdr(lo); | |
1163 | spin_unlock(&ino->i_lock); | |
1164 | pnfs_free_layout_hdr(lo); | |
1165 | } else | |
1166 | spin_unlock(&ino->i_lock); | |
f7e8917a FI |
1167 | } |
1168 | ||
1169 | void pnfs_roc_set_barrier(struct inode *ino, u32 barrier) | |
1170 | { | |
1171 | struct pnfs_layout_hdr *lo; | |
1172 | ||
1173 | spin_lock(&ino->i_lock); | |
1174 | lo = NFS_I(ino)->layout; | |
0654cc72 | 1175 | pnfs_mark_layout_returned_if_empty(lo); |
0f35ad6f | 1176 | if (pnfs_seqid_is_newer(barrier, lo->plh_barrier)) |
f7e8917a FI |
1177 | lo->plh_barrier = barrier; |
1178 | spin_unlock(&ino->i_lock); | |
6a463beb | 1179 | trace_nfs4_layoutreturn_on_close(ino, 0); |
f7e8917a FI |
1180 | } |
1181 | ||
4ff376fe | 1182 | void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier) |
f7e8917a FI |
1183 | { |
1184 | struct nfs_inode *nfsi = NFS_I(ino); | |
7fdab069 | 1185 | struct pnfs_layout_hdr *lo; |
7fdab069 | 1186 | u32 current_seqid; |
f7e8917a FI |
1187 | |
1188 | spin_lock(&ino->i_lock); | |
7fdab069 TM |
1189 | lo = nfsi->layout; |
1190 | current_seqid = be32_to_cpu(lo->plh_stateid.seqid); | |
f7e8917a | 1191 | |
7fdab069 TM |
1192 | /* Since close does not return a layout stateid for use as |
1193 | * a barrier, we choose the worst-case barrier. | |
1194 | */ | |
1195 | *barrier = current_seqid + atomic_read(&lo->plh_outstanding); | |
f7e8917a | 1196 | spin_unlock(&ino->i_lock); |
f7e8917a FI |
1197 | } |
1198 | ||
500d701f PT |
1199 | bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task) |
1200 | { | |
1201 | struct nfs_inode *nfsi = NFS_I(ino); | |
1202 | struct pnfs_layout_hdr *lo; | |
1203 | bool sleep = false; | |
1204 | ||
1205 | /* we might not have grabbed lo reference. so need to check under | |
1206 | * i_lock */ | |
1207 | spin_lock(&ino->i_lock); | |
1208 | lo = nfsi->layout; | |
1209 | if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) | |
1210 | sleep = true; | |
1211 | spin_unlock(&ino->i_lock); | |
1212 | ||
1213 | if (sleep) | |
1214 | rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL); | |
1215 | ||
1216 | return sleep; | |
1217 | } | |
1218 | ||
b1f69b75 AA |
1219 | /* |
1220 | * Compare two layout segments for sorting into layout cache. | |
1221 | * We want to preferentially return RW over RO layouts, so ensure those | |
1222 | * are seen first. | |
1223 | */ | |
1224 | static s64 | |
7dc0ac70 | 1225 | pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1, |
3cb2df17 | 1226 | const struct pnfs_layout_range *l2) |
b1f69b75 | 1227 | { |
fb3296eb BH |
1228 | s64 d; |
1229 | ||
1230 | /* high offset > low offset */ | |
1231 | d = l1->offset - l2->offset; | |
1232 | if (d) | |
1233 | return d; | |
1234 | ||
1235 | /* short length > long length */ | |
1236 | d = l2->length - l1->length; | |
1237 | if (d) | |
1238 | return d; | |
1239 | ||
b1f69b75 | 1240 | /* read > read/write */ |
fb3296eb | 1241 | return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ); |
b1f69b75 AA |
1242 | } |
1243 | ||
03772d2f TM |
1244 | static bool |
1245 | pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1, | |
1246 | const struct pnfs_layout_range *l2) | |
1247 | { | |
1248 | return pnfs_lseg_range_cmp(l1, l2) > 0; | |
1249 | } | |
1250 | ||
1251 | static bool | |
1252 | pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg, | |
1253 | struct pnfs_layout_segment *old) | |
1254 | { | |
1255 | return false; | |
1256 | } | |
1257 | ||
1258 | void | |
1259 | pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo, | |
1260 | struct pnfs_layout_segment *lseg, | |
1261 | bool (*is_after)(const struct pnfs_layout_range *, | |
1262 | const struct pnfs_layout_range *), | |
1263 | bool (*do_merge)(struct pnfs_layout_segment *, | |
1264 | struct pnfs_layout_segment *), | |
1265 | struct list_head *free_me) | |
974cec8c | 1266 | { |
03772d2f | 1267 | struct pnfs_layout_segment *lp, *tmp; |
b1f69b75 | 1268 | |
974cec8c AA |
1269 | dprintk("%s:Begin\n", __func__); |
1270 | ||
03772d2f TM |
1271 | list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) { |
1272 | if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0) | |
1273 | continue; | |
1274 | if (do_merge(lseg, lp)) { | |
1275 | mark_lseg_invalid(lp, free_me); | |
1276 | continue; | |
1277 | } | |
1278 | if (is_after(&lseg->pls_range, &lp->pls_range)) | |
b1f69b75 | 1279 | continue; |
566052c5 | 1280 | list_add_tail(&lseg->pls_list, &lp->pls_list); |
b1f69b75 AA |
1281 | dprintk("%s: inserted lseg %p " |
1282 | "iomode %d offset %llu length %llu before " | |
1283 | "lp %p iomode %d offset %llu length %llu\n", | |
566052c5 FI |
1284 | __func__, lseg, lseg->pls_range.iomode, |
1285 | lseg->pls_range.offset, lseg->pls_range.length, | |
1286 | lp, lp->pls_range.iomode, lp->pls_range.offset, | |
1287 | lp->pls_range.length); | |
fb3296eb | 1288 | goto out; |
974cec8c | 1289 | } |
fb3296eb BH |
1290 | list_add_tail(&lseg->pls_list, &lo->plh_segs); |
1291 | dprintk("%s: inserted lseg %p " | |
1292 | "iomode %d offset %llu length %llu at tail\n", | |
1293 | __func__, lseg, lseg->pls_range.iomode, | |
1294 | lseg->pls_range.offset, lseg->pls_range.length); | |
1295 | out: | |
70c3bd2b | 1296 | pnfs_get_layout_hdr(lo); |
974cec8c AA |
1297 | |
1298 | dprintk("%s:Return\n", __func__); | |
e5e94017 | 1299 | } |
03772d2f TM |
1300 | EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg); |
1301 | ||
1302 | static void | |
1303 | pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo, | |
1304 | struct pnfs_layout_segment *lseg, | |
1305 | struct list_head *free_me) | |
1306 | { | |
1307 | struct inode *inode = lo->plh_inode; | |
1308 | struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld; | |
1309 | ||
1310 | if (ld->add_lseg != NULL) | |
1311 | ld->add_lseg(lo, lseg, free_me); | |
1312 | else | |
1313 | pnfs_generic_layout_insert_lseg(lo, lseg, | |
1314 | pnfs_lseg_range_is_after, | |
1315 | pnfs_lseg_no_merge, | |
1316 | free_me); | |
1317 | } | |
e5e94017 BH |
1318 | |
1319 | static struct pnfs_layout_hdr * | |
9fa40758 PT |
1320 | alloc_init_layout_hdr(struct inode *ino, |
1321 | struct nfs_open_context *ctx, | |
1322 | gfp_t gfp_flags) | |
e5e94017 BH |
1323 | { |
1324 | struct pnfs_layout_hdr *lo; | |
1325 | ||
636fb9c8 | 1326 | lo = pnfs_alloc_layout_hdr(ino, gfp_flags); |
e5e94017 BH |
1327 | if (!lo) |
1328 | return NULL; | |
cc6e5340 | 1329 | atomic_set(&lo->plh_refcount, 1); |
b7edfaa1 FI |
1330 | INIT_LIST_HEAD(&lo->plh_layouts); |
1331 | INIT_LIST_HEAD(&lo->plh_segs); | |
fd9a8d71 | 1332 | INIT_LIST_HEAD(&lo->plh_bulk_destroy); |
b7edfaa1 | 1333 | lo->plh_inode = ino; |
5cc2216d | 1334 | lo->plh_lc_cred = get_rpccred(ctx->cred); |
e5e94017 BH |
1335 | return lo; |
1336 | } | |
1337 | ||
1338 | static struct pnfs_layout_hdr * | |
9fa40758 PT |
1339 | pnfs_find_alloc_layout(struct inode *ino, |
1340 | struct nfs_open_context *ctx, | |
1341 | gfp_t gfp_flags) | |
e5e94017 BH |
1342 | { |
1343 | struct nfs_inode *nfsi = NFS_I(ino); | |
1344 | struct pnfs_layout_hdr *new = NULL; | |
1345 | ||
1346 | dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout); | |
1347 | ||
251ec410 TM |
1348 | if (nfsi->layout != NULL) |
1349 | goto out_existing; | |
e5e94017 | 1350 | spin_unlock(&ino->i_lock); |
9fa40758 | 1351 | new = alloc_init_layout_hdr(ino, ctx, gfp_flags); |
e5e94017 BH |
1352 | spin_lock(&ino->i_lock); |
1353 | ||
251ec410 | 1354 | if (likely(nfsi->layout == NULL)) { /* Won the race? */ |
e5e94017 | 1355 | nfsi->layout = new; |
251ec410 | 1356 | return new; |
7175fe90 YN |
1357 | } else if (new != NULL) |
1358 | pnfs_free_layout_hdr(new); | |
251ec410 TM |
1359 | out_existing: |
1360 | pnfs_get_layout_hdr(nfsi->layout); | |
e5e94017 BH |
1361 | return nfsi->layout; |
1362 | } | |
1363 | ||
b1f69b75 AA |
1364 | /* |
1365 | * iomode matching rules: | |
1366 | * iomode lseg match | |
1367 | * ----- ----- ----- | |
1368 | * ANY READ true | |
1369 | * ANY RW true | |
1370 | * RW READ false | |
1371 | * RW RW true | |
1372 | * READ READ true | |
1373 | * READ RW true | |
1374 | */ | |
3cb2df17 | 1375 | static bool |
7dc0ac70 | 1376 | pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range, |
3cb2df17 | 1377 | const struct pnfs_layout_range *range) |
b1f69b75 | 1378 | { |
fb3296eb BH |
1379 | struct pnfs_layout_range range1; |
1380 | ||
1381 | if ((range->iomode == IOMODE_RW && | |
1382 | ls_range->iomode != IOMODE_RW) || | |
7dc0ac70 | 1383 | !pnfs_lseg_range_intersecting(ls_range, range)) |
fb3296eb BH |
1384 | return 0; |
1385 | ||
1386 | /* range1 covers only the first byte in the range */ | |
1387 | range1 = *range; | |
1388 | range1.length = 1; | |
7dc0ac70 | 1389 | return pnfs_lseg_range_contained(ls_range, &range1); |
b1f69b75 AA |
1390 | } |
1391 | ||
1392 | /* | |
1393 | * lookup range in layout | |
1394 | */ | |
e5e94017 | 1395 | static struct pnfs_layout_segment * |
fb3296eb BH |
1396 | pnfs_find_lseg(struct pnfs_layout_hdr *lo, |
1397 | struct pnfs_layout_range *range) | |
e5e94017 | 1398 | { |
b1f69b75 AA |
1399 | struct pnfs_layout_segment *lseg, *ret = NULL; |
1400 | ||
1401 | dprintk("%s:Begin\n", __func__); | |
1402 | ||
b7edfaa1 | 1403 | list_for_each_entry(lseg, &lo->plh_segs, pls_list) { |
4541d16c | 1404 | if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) && |
ce6ab4f2 | 1405 | !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) && |
7dc0ac70 | 1406 | pnfs_lseg_range_match(&lseg->pls_range, range)) { |
9369a431 | 1407 | ret = pnfs_get_lseg(lseg); |
b1f69b75 AA |
1408 | break; |
1409 | } | |
b1f69b75 AA |
1410 | } |
1411 | ||
1412 | dprintk("%s:Return lseg %p ref %d\n", | |
4541d16c | 1413 | __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0); |
b1f69b75 | 1414 | return ret; |
e5e94017 BH |
1415 | } |
1416 | ||
d23d61c8 AA |
1417 | /* |
1418 | * Use mdsthreshold hints set at each OPEN to determine if I/O should go | |
1419 | * to the MDS or over pNFS | |
1420 | * | |
1421 | * The nfs_inode read_io and write_io fields are cumulative counters reset | |
1422 | * when there are no layout segments. Note that in pnfs_update_layout iomode | |
1423 | * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a | |
1424 | * WRITE request. | |
1425 | * | |
1426 | * A return of true means use MDS I/O. | |
1427 | * | |
1428 | * From rfc 5661: | |
1429 | * If a file's size is smaller than the file size threshold, data accesses | |
1430 | * SHOULD be sent to the metadata server. If an I/O request has a length that | |
1431 | * is below the I/O size threshold, the I/O SHOULD be sent to the metadata | |
1432 | * server. If both file size and I/O size are provided, the client SHOULD | |
1433 | * reach or exceed both thresholds before sending its read or write | |
1434 | * requests to the data server. | |
1435 | */ | |
1436 | static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx, | |
1437 | struct inode *ino, int iomode) | |
1438 | { | |
1439 | struct nfs4_threshold *t = ctx->mdsthreshold; | |
1440 | struct nfs_inode *nfsi = NFS_I(ino); | |
1441 | loff_t fsize = i_size_read(ino); | |
1442 | bool size = false, size_set = false, io = false, io_set = false, ret = false; | |
1443 | ||
1444 | if (t == NULL) | |
1445 | return ret; | |
1446 | ||
1447 | dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n", | |
1448 | __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz); | |
1449 | ||
1450 | switch (iomode) { | |
1451 | case IOMODE_READ: | |
1452 | if (t->bm & THRESHOLD_RD) { | |
1453 | dprintk("%s fsize %llu\n", __func__, fsize); | |
1454 | size_set = true; | |
1455 | if (fsize < t->rd_sz) | |
1456 | size = true; | |
1457 | } | |
1458 | if (t->bm & THRESHOLD_RD_IO) { | |
1459 | dprintk("%s nfsi->read_io %llu\n", __func__, | |
1460 | nfsi->read_io); | |
1461 | io_set = true; | |
1462 | if (nfsi->read_io < t->rd_io_sz) | |
1463 | io = true; | |
1464 | } | |
1465 | break; | |
1466 | case IOMODE_RW: | |
1467 | if (t->bm & THRESHOLD_WR) { | |
1468 | dprintk("%s fsize %llu\n", __func__, fsize); | |
1469 | size_set = true; | |
1470 | if (fsize < t->wr_sz) | |
1471 | size = true; | |
1472 | } | |
1473 | if (t->bm & THRESHOLD_WR_IO) { | |
1474 | dprintk("%s nfsi->write_io %llu\n", __func__, | |
1475 | nfsi->write_io); | |
1476 | io_set = true; | |
1477 | if (nfsi->write_io < t->wr_io_sz) | |
1478 | io = true; | |
1479 | } | |
1480 | break; | |
1481 | } | |
1482 | if (size_set && io_set) { | |
1483 | if (size && io) | |
1484 | ret = true; | |
1485 | } else if (size || io) | |
1486 | ret = true; | |
1487 | ||
1488 | dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret); | |
1489 | return ret; | |
1490 | } | |
1491 | ||
aa8a45ee PT |
1492 | static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo) |
1493 | { | |
1494 | /* | |
1495 | * send layoutcommit as it can hold up layoutreturn due to lseg | |
1496 | * reference | |
1497 | */ | |
1498 | pnfs_layoutcommit_inode(lo->plh_inode, false); | |
1499 | return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN, | |
2e5b29f0 | 1500 | nfs_wait_bit_killable, |
aa8a45ee PT |
1501 | TASK_UNINTERRUPTIBLE); |
1502 | } | |
1503 | ||
d67ae825 TH |
1504 | static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo) |
1505 | { | |
1506 | unsigned long *bitlock = &lo->plh_flags; | |
1507 | ||
1508 | clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock); | |
1509 | smp_mb__after_atomic(); | |
1510 | wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET); | |
1511 | } | |
1512 | ||
e5e94017 BH |
1513 | /* |
1514 | * Layout segment is retreived from the server if not cached. | |
1515 | * The appropriate layout segment is referenced and returned to the caller. | |
1516 | */ | |
7c24d948 | 1517 | struct pnfs_layout_segment * |
e5e94017 BH |
1518 | pnfs_update_layout(struct inode *ino, |
1519 | struct nfs_open_context *ctx, | |
fb3296eb BH |
1520 | loff_t pos, |
1521 | u64 count, | |
a75b9df9 TM |
1522 | enum pnfs_iomode iomode, |
1523 | gfp_t gfp_flags) | |
e5e94017 | 1524 | { |
fb3296eb BH |
1525 | struct pnfs_layout_range arg = { |
1526 | .iomode = iomode, | |
1527 | .offset = pos, | |
1528 | .length = count, | |
1529 | }; | |
707ed5fd | 1530 | unsigned pg_offset; |
6382a441 WAA |
1531 | struct nfs_server *server = NFS_SERVER(ino); |
1532 | struct nfs_client *clp = server->nfs_client; | |
e5e94017 BH |
1533 | struct pnfs_layout_hdr *lo; |
1534 | struct pnfs_layout_segment *lseg = NULL; | |
30005121 | 1535 | bool first; |
e5e94017 | 1536 | |
9a4bf31d | 1537 | if (!pnfs_enabled_sb(NFS_SERVER(ino))) { |
f4848303 | 1538 | trace_pnfs_update_layout(ino, pos, count, iomode, NULL, |
9a4bf31d | 1539 | PNFS_UPDATE_LAYOUT_NO_PNFS); |
f86bbcf8 | 1540 | goto out; |
9a4bf31d | 1541 | } |
d23d61c8 | 1542 | |
9a4bf31d | 1543 | if (iomode == IOMODE_READ && i_size_read(ino) == 0) { |
f4848303 | 1544 | trace_pnfs_update_layout(ino, pos, count, iomode, NULL, |
9a4bf31d | 1545 | PNFS_UPDATE_LAYOUT_RD_ZEROLEN); |
4ae93560 | 1546 | goto out; |
9a4bf31d | 1547 | } |
4ae93560 | 1548 | |
9a4bf31d | 1549 | if (pnfs_within_mdsthreshold(ctx, ino, iomode)) { |
f4848303 | 1550 | trace_pnfs_update_layout(ino, pos, count, iomode, NULL, |
9a4bf31d | 1551 | PNFS_UPDATE_LAYOUT_MDSTHRESH); |
f86bbcf8 | 1552 | goto out; |
9a4bf31d | 1553 | } |
d23d61c8 | 1554 | |
9bf87482 PT |
1555 | lookup_again: |
1556 | first = false; | |
e5e94017 | 1557 | spin_lock(&ino->i_lock); |
9fa40758 | 1558 | lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags); |
830ffb56 TM |
1559 | if (lo == NULL) { |
1560 | spin_unlock(&ino->i_lock); | |
f4848303 | 1561 | trace_pnfs_update_layout(ino, pos, count, iomode, NULL, |
9a4bf31d | 1562 | PNFS_UPDATE_LAYOUT_NOMEM); |
830ffb56 TM |
1563 | goto out; |
1564 | } | |
e5e94017 | 1565 | |
43f1b3da | 1566 | /* Do we even need to bother with this? */ |
a59c30ac | 1567 | if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) { |
f4848303 | 1568 | trace_pnfs_update_layout(ino, pos, count, iomode, lo, |
9a4bf31d | 1569 | PNFS_UPDATE_LAYOUT_BULK_RECALL); |
43f1b3da | 1570 | dprintk("%s matches recall, use MDS\n", __func__); |
e5e94017 BH |
1571 | goto out_unlock; |
1572 | } | |
1573 | ||
1574 | /* if LAYOUTGET already failed once we don't try again */ | |
2e5b29f0 | 1575 | if (pnfs_layout_io_test_failed(lo, iomode)) { |
f4848303 | 1576 | trace_pnfs_update_layout(ino, pos, count, iomode, lo, |
9a4bf31d | 1577 | PNFS_UPDATE_LAYOUT_IO_TEST_FAIL); |
e5e94017 | 1578 | goto out_unlock; |
9a4bf31d | 1579 | } |
e5e94017 | 1580 | |
9bf87482 PT |
1581 | first = list_empty(&lo->plh_segs); |
1582 | if (first) { | |
1583 | /* The first layoutget for the file. Need to serialize per | |
1584 | * RFC 5661 Errata 3208. | |
1585 | */ | |
1586 | if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, | |
1587 | &lo->plh_flags)) { | |
1588 | spin_unlock(&ino->i_lock); | |
1589 | wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET, | |
1590 | TASK_UNINTERRUPTIBLE); | |
1591 | pnfs_put_layout_hdr(lo); | |
1592 | goto lookup_again; | |
1593 | } | |
1594 | } else { | |
1595 | /* Check to see if the layout for the given range | |
1596 | * already exists | |
1597 | */ | |
1598 | lseg = pnfs_find_lseg(lo, &arg); | |
9a4bf31d | 1599 | if (lseg) { |
f4848303 | 1600 | trace_pnfs_update_layout(ino, pos, count, iomode, lo, |
9a4bf31d | 1601 | PNFS_UPDATE_LAYOUT_FOUND_CACHED); |
9bf87482 | 1602 | goto out_unlock; |
9a4bf31d | 1603 | } |
9bf87482 | 1604 | } |
568e8c49 | 1605 | |
aa8a45ee PT |
1606 | /* |
1607 | * Because we free lsegs before sending LAYOUTRETURN, we need to wait | |
1608 | * for LAYOUTRETURN even if first is true. | |
1609 | */ | |
8f70f53a | 1610 | if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) { |
aa8a45ee PT |
1611 | spin_unlock(&ino->i_lock); |
1612 | dprintk("%s wait for layoutreturn\n", __func__); | |
1613 | if (pnfs_prepare_to_retry_layoutget(lo)) { | |
d67ae825 TH |
1614 | if (first) |
1615 | pnfs_clear_first_layoutget(lo); | |
aa8a45ee PT |
1616 | pnfs_put_layout_hdr(lo); |
1617 | dprintk("%s retrying\n", __func__); | |
1618 | goto lookup_again; | |
1619 | } | |
f4848303 | 1620 | trace_pnfs_update_layout(ino, pos, count, iomode, lo, |
9a4bf31d | 1621 | PNFS_UPDATE_LAYOUT_RETURN); |
aa8a45ee PT |
1622 | goto out_put_layout_hdr; |
1623 | } | |
1624 | ||
9a4bf31d | 1625 | if (pnfs_layoutgets_blocked(lo)) { |
f4848303 | 1626 | trace_pnfs_update_layout(ino, pos, count, iomode, lo, |
9a4bf31d | 1627 | PNFS_UPDATE_LAYOUT_BLOCKED); |
cf7d63f1 | 1628 | goto out_unlock; |
9a4bf31d | 1629 | } |
cf7d63f1 | 1630 | atomic_inc(&lo->plh_outstanding); |
f49f9baa | 1631 | spin_unlock(&ino->i_lock); |
30005121 | 1632 | |
abb9a007 | 1633 | if (list_empty(&lo->plh_layouts)) { |
2130ff66 FI |
1634 | /* The lo must be on the clp list if there is any |
1635 | * chance of a CB_LAYOUTRECALL(FILE) coming in. | |
1636 | */ | |
1637 | spin_lock(&clp->cl_lock); | |
abb9a007 PT |
1638 | if (list_empty(&lo->plh_layouts)) |
1639 | list_add_tail(&lo->plh_layouts, &server->layouts); | |
2130ff66 FI |
1640 | spin_unlock(&clp->cl_lock); |
1641 | } | |
e5e94017 | 1642 | |
09cbfeaf | 1643 | pg_offset = arg.offset & ~PAGE_MASK; |
707ed5fd BH |
1644 | if (pg_offset) { |
1645 | arg.offset -= pg_offset; | |
1646 | arg.length += pg_offset; | |
1647 | } | |
7c24d948 | 1648 | if (arg.length != NFS4_MAX_UINT64) |
09cbfeaf | 1649 | arg.length = PAGE_ALIGN(arg.length); |
707ed5fd | 1650 | |
fb3296eb | 1651 | lseg = send_layoutget(lo, ctx, &arg, gfp_flags); |
cf7d63f1 | 1652 | atomic_dec(&lo->plh_outstanding); |
f4848303 | 1653 | trace_pnfs_update_layout(ino, pos, count, iomode, lo, |
9a4bf31d | 1654 | PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET); |
830ffb56 | 1655 | out_put_layout_hdr: |
d67ae825 TH |
1656 | if (first) |
1657 | pnfs_clear_first_layoutget(lo); | |
70c3bd2b | 1658 | pnfs_put_layout_hdr(lo); |
e5e94017 | 1659 | out: |
f86bbcf8 TM |
1660 | dprintk("%s: inode %s/%llu pNFS layout segment %s for " |
1661 | "(%s, offset: %llu, length: %llu)\n", | |
1662 | __func__, ino->i_sb->s_id, | |
1663 | (unsigned long long)NFS_FILEID(ino), | |
d600ad1f | 1664 | IS_ERR_OR_NULL(lseg) ? "not found" : "found", |
f86bbcf8 TM |
1665 | iomode==IOMODE_RW ? "read/write" : "read-only", |
1666 | (unsigned long long)pos, | |
1667 | (unsigned long long)count); | |
e5e94017 BH |
1668 | return lseg; |
1669 | out_unlock: | |
1670 | spin_unlock(&ino->i_lock); | |
830ffb56 | 1671 | goto out_put_layout_hdr; |
e5e94017 | 1672 | } |
7c24d948 | 1673 | EXPORT_SYMBOL_GPL(pnfs_update_layout); |
b1f69b75 | 1674 | |
540d9864 TM |
1675 | static bool |
1676 | pnfs_sanity_check_layout_range(struct pnfs_layout_range *range) | |
1677 | { | |
1678 | switch (range->iomode) { | |
1679 | case IOMODE_READ: | |
1680 | case IOMODE_RW: | |
1681 | break; | |
1682 | default: | |
1683 | return false; | |
1684 | } | |
1685 | if (range->offset == NFS4_MAX_UINT64) | |
1686 | return false; | |
1687 | if (range->length == 0) | |
1688 | return false; | |
1689 | if (range->length != NFS4_MAX_UINT64 && | |
1690 | range->length > NFS4_MAX_UINT64 - range->offset) | |
1691 | return false; | |
1692 | return true; | |
1693 | } | |
1694 | ||
a0b0a6e3 | 1695 | struct pnfs_layout_segment * |
b1f69b75 AA |
1696 | pnfs_layout_process(struct nfs4_layoutget *lgp) |
1697 | { | |
1698 | struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout; | |
1699 | struct nfs4_layoutget_res *res = &lgp->res; | |
1700 | struct pnfs_layout_segment *lseg; | |
b7edfaa1 | 1701 | struct inode *ino = lo->plh_inode; |
78096cca | 1702 | LIST_HEAD(free_me); |
540d9864 TM |
1703 | int status = -EINVAL; |
1704 | ||
1705 | if (!pnfs_sanity_check_layout_range(&res->range)) | |
1706 | goto out; | |
b1f69b75 AA |
1707 | |
1708 | /* Inject layout blob into I/O device driver */ | |
a75b9df9 | 1709 | lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags); |
b1f69b75 AA |
1710 | if (!lseg || IS_ERR(lseg)) { |
1711 | if (!lseg) | |
1712 | status = -ENOMEM; | |
1713 | else | |
1714 | status = PTR_ERR(lseg); | |
1715 | dprintk("%s: Could not allocate layout: error %d\n", | |
1716 | __func__, status); | |
1717 | goto out; | |
1718 | } | |
1719 | ||
1013df61 CH |
1720 | init_lseg(lo, lseg); |
1721 | lseg->pls_range = res->range; | |
66755283 | 1722 | lseg->pls_seq = be32_to_cpu(res->stateid.seqid); |
1013df61 | 1723 | |
b1f69b75 | 1724 | spin_lock(&ino->i_lock); |
e1c06f80 | 1725 | if (pnfs_layoutgets_blocked(lo)) { |
43f1b3da FI |
1726 | dprintk("%s forget reply due to state\n", __func__); |
1727 | goto out_forget_reply; | |
1728 | } | |
038d6493 | 1729 | |
362f7474 CH |
1730 | if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) { |
1731 | /* existing state ID, make sure the sequence number matches. */ | |
1732 | if (pnfs_layout_stateid_blocked(lo, &res->stateid)) { | |
1733 | dprintk("%s forget reply due to sequence\n", __func__); | |
4f2e9dce | 1734 | status = -EAGAIN; |
362f7474 CH |
1735 | goto out_forget_reply; |
1736 | } | |
1737 | pnfs_set_layout_stateid(lo, &res->stateid, false); | |
1738 | } else { | |
1739 | /* | |
1740 | * We got an entirely new state ID. Mark all segments for the | |
1741 | * inode invalid, and don't bother validating the stateid | |
1742 | * sequence number. | |
1743 | */ | |
6d597e17 | 1744 | pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL, 0); |
362f7474 CH |
1745 | |
1746 | nfs4_stateid_copy(&lo->plh_stateid, &res->stateid); | |
1747 | lo->plh_barrier = be32_to_cpu(res->stateid.seqid); | |
1748 | } | |
038d6493 | 1749 | |
47abadef CH |
1750 | clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags); |
1751 | ||
9369a431 | 1752 | pnfs_get_lseg(lseg); |
03772d2f | 1753 | pnfs_layout_insert_lseg(lo, lseg, &free_me); |
b1f69b75 | 1754 | |
3976143b | 1755 | if (res->return_on_close) |
f7e8917a | 1756 | set_bit(NFS_LSEG_ROC, &lseg->pls_flags); |
f7e8917a | 1757 | |
b1f69b75 | 1758 | spin_unlock(&ino->i_lock); |
78096cca | 1759 | pnfs_free_lseg_list(&free_me); |
a0b0a6e3 | 1760 | return lseg; |
b1f69b75 | 1761 | out: |
a0b0a6e3 | 1762 | return ERR_PTR(status); |
43f1b3da FI |
1763 | |
1764 | out_forget_reply: | |
1765 | spin_unlock(&ino->i_lock); | |
1766 | lseg->pls_layout = lo; | |
1767 | NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg); | |
1768 | goto out; | |
b1f69b75 AA |
1769 | } |
1770 | ||
016256df | 1771 | static void |
3982a6a2 JL |
1772 | pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode, |
1773 | u32 seq) | |
5c97f5de TM |
1774 | { |
1775 | if (lo->plh_return_iomode == iomode) | |
1776 | return; | |
1777 | if (lo->plh_return_iomode != 0) | |
1778 | iomode = IOMODE_ANY; | |
1779 | lo->plh_return_iomode = iomode; | |
e0fa0d01 | 1780 | set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags); |
3982a6a2 JL |
1781 | if (!lo->plh_return_seq || pnfs_seqid_is_newer(seq, lo->plh_return_seq)) |
1782 | lo->plh_return_seq = seq; | |
5c97f5de TM |
1783 | } |
1784 | ||
2f215968 TM |
1785 | /** |
1786 | * pnfs_mark_matching_lsegs_return - Free or return matching layout segments | |
1787 | * @lo: pointer to layout header | |
1788 | * @tmp_list: list header to be used with pnfs_free_lseg_list() | |
1789 | * @return_range: describe layout segment ranges to be returned | |
1790 | * | |
1791 | * This function is mainly intended for use by layoutrecall. It attempts | |
1792 | * to free the layout segment immediately, or else to mark it for return | |
1793 | * as soon as its reference count drops to zero. | |
1794 | */ | |
10335556 | 1795 | int |
016256df PT |
1796 | pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo, |
1797 | struct list_head *tmp_list, | |
6d597e17 JL |
1798 | const struct pnfs_layout_range *return_range, |
1799 | u32 seq) | |
016256df PT |
1800 | { |
1801 | struct pnfs_layout_segment *lseg, *next; | |
10335556 | 1802 | int remaining = 0; |
016256df PT |
1803 | |
1804 | dprintk("%s:Begin lo %p\n", __func__, lo); | |
1805 | ||
1806 | if (list_empty(&lo->plh_segs)) | |
10335556 | 1807 | return 0; |
016256df | 1808 | |
fc7ff367 | 1809 | assert_spin_locked(&lo->plh_inode->i_lock); |
016256df PT |
1810 | |
1811 | list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) | |
1812 | if (should_free_lseg(&lseg->pls_range, return_range)) { | |
1813 | dprintk("%s: marking lseg %p iomode %d " | |
1814 | "offset %llu length %llu\n", __func__, | |
1815 | lseg, lseg->pls_range.iomode, | |
1816 | lseg->pls_range.offset, | |
1817 | lseg->pls_range.length); | |
2f215968 TM |
1818 | if (mark_lseg_invalid(lseg, tmp_list)) |
1819 | continue; | |
1820 | remaining++; | |
016256df | 1821 | set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags); |
016256df | 1822 | } |
6d597e17 JL |
1823 | |
1824 | if (remaining) | |
1825 | pnfs_set_plh_return_info(lo, return_range->iomode, seq); | |
1826 | ||
10335556 | 1827 | return remaining; |
016256df PT |
1828 | } |
1829 | ||
1830 | void pnfs_error_mark_layout_for_return(struct inode *inode, | |
1831 | struct pnfs_layout_segment *lseg) | |
1832 | { | |
1833 | struct pnfs_layout_hdr *lo = NFS_I(inode)->layout; | |
016256df PT |
1834 | struct pnfs_layout_range range = { |
1835 | .iomode = lseg->pls_range.iomode, | |
1836 | .offset = 0, | |
1837 | .length = NFS4_MAX_UINT64, | |
1838 | }; | |
1839 | LIST_HEAD(free_me); | |
10335556 | 1840 | bool return_now = false; |
016256df PT |
1841 | |
1842 | spin_lock(&inode->i_lock); | |
3982a6a2 | 1843 | pnfs_set_plh_return_info(lo, range.iomode, lseg->pls_seq); |
016256df PT |
1844 | /* |
1845 | * mark all matching lsegs so that we are sure to have no live | |
1846 | * segments at hand when sending layoutreturn. See pnfs_put_lseg() | |
1847 | * for how it works. | |
1848 | */ | |
6d597e17 JL |
1849 | if (!pnfs_mark_matching_lsegs_return(lo, &free_me, |
1850 | &range, lseg->pls_seq)) { | |
10335556 TM |
1851 | nfs4_stateid stateid; |
1852 | enum pnfs_iomode iomode = lo->plh_return_iomode; | |
1853 | ||
1854 | nfs4_stateid_copy(&stateid, &lo->plh_stateid); | |
1855 | return_now = pnfs_prepare_layoutreturn(lo); | |
1856 | spin_unlock(&inode->i_lock); | |
1857 | if (return_now) | |
1858 | pnfs_send_layoutreturn(lo, &stateid, iomode, false); | |
1859 | } else { | |
1860 | spin_unlock(&inode->i_lock); | |
1861 | nfs_commit_inode(inode, 0); | |
1862 | } | |
016256df PT |
1863 | pnfs_free_lseg_list(&free_me); |
1864 | } | |
1865 | EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return); | |
1866 | ||
d8007d4d TM |
1867 | void |
1868 | pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req) | |
1869 | { | |
1fd937bd PT |
1870 | u64 rd_size = req->wb_bytes; |
1871 | ||
cb5d04bc PT |
1872 | if (pgio->pg_lseg == NULL) { |
1873 | if (pgio->pg_dreq == NULL) | |
1874 | rd_size = i_size_read(pgio->pg_inode) - req_offset(req); | |
1875 | else | |
1876 | rd_size = nfs_dreq_bytes_left(pgio->pg_dreq); | |
1877 | ||
1878 | pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, | |
1879 | req->wb_context, | |
1880 | req_offset(req), | |
1881 | rd_size, | |
1882 | IOMODE_READ, | |
1883 | GFP_KERNEL); | |
d600ad1f PT |
1884 | if (IS_ERR(pgio->pg_lseg)) { |
1885 | pgio->pg_error = PTR_ERR(pgio->pg_lseg); | |
1886 | pgio->pg_lseg = NULL; | |
1887 | return; | |
1888 | } | |
cb5d04bc | 1889 | } |
e885de1a TM |
1890 | /* If no lseg, fall back to read through mds */ |
1891 | if (pgio->pg_lseg == NULL) | |
1f945357 | 1892 | nfs_pageio_reset_read_mds(pgio); |
e885de1a | 1893 | |
d8007d4d TM |
1894 | } |
1895 | EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read); | |
1896 | ||
1897 | void | |
6296556f PT |
1898 | pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, |
1899 | struct nfs_page *req, u64 wb_size) | |
d8007d4d | 1900 | { |
d600ad1f | 1901 | if (pgio->pg_lseg == NULL) { |
cb5d04bc PT |
1902 | pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, |
1903 | req->wb_context, | |
1904 | req_offset(req), | |
1905 | wb_size, | |
1906 | IOMODE_RW, | |
1907 | GFP_NOFS); | |
d600ad1f PT |
1908 | if (IS_ERR(pgio->pg_lseg)) { |
1909 | pgio->pg_error = PTR_ERR(pgio->pg_lseg); | |
1910 | pgio->pg_lseg = NULL; | |
1911 | return; | |
1912 | } | |
1913 | } | |
e885de1a TM |
1914 | /* If no lseg, fall back to write through mds */ |
1915 | if (pgio->pg_lseg == NULL) | |
1f945357 | 1916 | nfs_pageio_reset_write_mds(pgio); |
d8007d4d TM |
1917 | } |
1918 | EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write); | |
1919 | ||
180bb5ec WAA |
1920 | void |
1921 | pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc) | |
1922 | { | |
1923 | if (desc->pg_lseg) { | |
1924 | pnfs_put_lseg(desc->pg_lseg); | |
1925 | desc->pg_lseg = NULL; | |
1926 | } | |
1927 | } | |
1928 | EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup); | |
1929 | ||
b4fdac1a WAA |
1930 | /* |
1931 | * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number | |
1932 | * of bytes (maximum @req->wb_bytes) that can be coalesced. | |
1933 | */ | |
1934 | size_t | |
a7d42ddb WAA |
1935 | pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, |
1936 | struct nfs_page *prev, struct nfs_page *req) | |
94ad1c80 | 1937 | { |
0f9c429e | 1938 | unsigned int size; |
c5e20cb7 | 1939 | u64 seg_end, req_start, seg_left; |
0f9c429e WAA |
1940 | |
1941 | size = nfs_generic_pg_test(pgio, prev, req); | |
0f9c429e WAA |
1942 | if (!size) |
1943 | return 0; | |
94ad1c80 | 1944 | |
19982ba8 | 1945 | /* |
c5e20cb7 WAA |
1946 | * 'size' contains the number of bytes left in the current page (up |
1947 | * to the original size asked for in @req->wb_bytes). | |
1948 | * | |
1949 | * Calculate how many bytes are left in the layout segment | |
1950 | * and if there are less bytes than 'size', return that instead. | |
19982ba8 TM |
1951 | * |
1952 | * Please also note that 'end_offset' is actually the offset of the | |
1953 | * first byte that lies outside the pnfs_layout_range. FIXME? | |
1954 | * | |
1955 | */ | |
19b54848 | 1956 | if (pgio->pg_lseg) { |
c5e20cb7 WAA |
1957 | seg_end = end_offset(pgio->pg_lseg->pls_range.offset, |
1958 | pgio->pg_lseg->pls_range.length); | |
1959 | req_start = req_offset(req); | |
7c13789e | 1960 | WARN_ON_ONCE(req_start >= seg_end); |
c5e20cb7 | 1961 | /* start of request is past the last byte of this segment */ |
7c13789e WAA |
1962 | if (req_start >= seg_end) { |
1963 | /* reference the new lseg */ | |
1964 | if (pgio->pg_ops->pg_cleanup) | |
1965 | pgio->pg_ops->pg_cleanup(pgio); | |
1966 | if (pgio->pg_ops->pg_init) | |
1967 | pgio->pg_ops->pg_init(pgio, req); | |
19b54848 | 1968 | return 0; |
7c13789e | 1969 | } |
c5e20cb7 WAA |
1970 | |
1971 | /* adjust 'size' iff there are fewer bytes left in the | |
1972 | * segment than what nfs_generic_pg_test returned */ | |
1973 | seg_left = seg_end - req_start; | |
1974 | if (seg_left < size) | |
1975 | size = (unsigned int)seg_left; | |
19b54848 | 1976 | } |
0f9c429e | 1977 | |
19b54848 | 1978 | return size; |
94ad1c80 | 1979 | } |
89a58e32 | 1980 | EXPORT_SYMBOL_GPL(pnfs_generic_pg_test); |
94ad1c80 | 1981 | |
53113ad3 | 1982 | int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr) |
e2fecb21 TM |
1983 | { |
1984 | struct nfs_pageio_descriptor pgio; | |
e2fecb21 TM |
1985 | |
1986 | /* Resend all requests through the MDS */ | |
53113ad3 WAA |
1987 | nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true, |
1988 | hdr->completion_ops); | |
c7070113 | 1989 | set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags); |
53113ad3 | 1990 | return nfs_pageio_resend(&pgio, hdr); |
e2fecb21 | 1991 | } |
e7dd79af | 1992 | EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds); |
e2fecb21 | 1993 | |
d45f60c6 | 1994 | static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr) |
1acbbb4e | 1995 | { |
cd841605 FI |
1996 | |
1997 | dprintk("pnfs write error = %d\n", hdr->pnfs_error); | |
1998 | if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags & | |
1acbbb4e | 1999 | PNFS_LAYOUTRET_ON_ERROR) { |
cd841605 | 2000 | pnfs_return_layout(hdr->inode); |
1acbbb4e | 2001 | } |
6c75dc0d | 2002 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) |
53113ad3 | 2003 | hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr); |
1acbbb4e FI |
2004 | } |
2005 | ||
d20581aa BH |
2006 | /* |
2007 | * Called by non rpc-based layout drivers | |
2008 | */ | |
d45f60c6 | 2009 | void pnfs_ld_write_done(struct nfs_pgio_header *hdr) |
44b83799 | 2010 | { |
f8417b48 | 2011 | if (likely(!hdr->pnfs_error)) { |
67af7611 TM |
2012 | pnfs_set_layoutcommit(hdr->inode, hdr->lseg, |
2013 | hdr->mds_offset + hdr->res.count); | |
d45f60c6 | 2014 | hdr->mds_ops->rpc_call_done(&hdr->task, hdr); |
f8417b48 KM |
2015 | } |
2016 | trace_nfs4_pnfs_write(hdr, hdr->pnfs_error); | |
2017 | if (unlikely(hdr->pnfs_error)) | |
d45f60c6 WAA |
2018 | pnfs_ld_handle_write_error(hdr); |
2019 | hdr->mds_ops->rpc_release(hdr); | |
44b83799 | 2020 | } |
d20581aa | 2021 | EXPORT_SYMBOL_GPL(pnfs_ld_write_done); |
44b83799 | 2022 | |
dce81290 TM |
2023 | static void |
2024 | pnfs_write_through_mds(struct nfs_pageio_descriptor *desc, | |
d45f60c6 | 2025 | struct nfs_pgio_header *hdr) |
dce81290 | 2026 | { |
48d635f1 | 2027 | struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc); |
a7d42ddb | 2028 | |
6c75dc0d | 2029 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { |
a7d42ddb | 2030 | list_splice_tail_init(&hdr->pages, &mirror->pg_list); |
6c75dc0d | 2031 | nfs_pageio_reset_write_mds(desc); |
a7d42ddb | 2032 | mirror->pg_recoalesce = 1; |
6c75dc0d | 2033 | } |
d45f60c6 | 2034 | nfs_pgio_data_destroy(hdr); |
1ca018d2 | 2035 | hdr->release(hdr); |
dce81290 TM |
2036 | } |
2037 | ||
2038 | static enum pnfs_try_status | |
d45f60c6 | 2039 | pnfs_try_to_write_data(struct nfs_pgio_header *hdr, |
dce81290 TM |
2040 | const struct rpc_call_ops *call_ops, |
2041 | struct pnfs_layout_segment *lseg, | |
2042 | int how) | |
0382b744 | 2043 | { |
cd841605 | 2044 | struct inode *inode = hdr->inode; |
0382b744 AA |
2045 | enum pnfs_try_status trypnfs; |
2046 | struct nfs_server *nfss = NFS_SERVER(inode); | |
2047 | ||
cd841605 | 2048 | hdr->mds_ops = call_ops; |
0382b744 AA |
2049 | |
2050 | dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__, | |
d45f60c6 WAA |
2051 | inode->i_ino, hdr->args.count, hdr->args.offset, how); |
2052 | trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how); | |
6c75dc0d | 2053 | if (trypnfs != PNFS_NOT_ATTEMPTED) |
0382b744 | 2054 | nfs_inc_stats(inode, NFSIOS_PNFS_WRITE); |
0382b744 AA |
2055 | dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs); |
2056 | return trypnfs; | |
2057 | } | |
2058 | ||
dce81290 | 2059 | static void |
7f714720 WAA |
2060 | pnfs_do_write(struct nfs_pageio_descriptor *desc, |
2061 | struct nfs_pgio_header *hdr, int how) | |
dce81290 | 2062 | { |
dce81290 TM |
2063 | const struct rpc_call_ops *call_ops = desc->pg_rpc_callops; |
2064 | struct pnfs_layout_segment *lseg = desc->pg_lseg; | |
7f714720 | 2065 | enum pnfs_try_status trypnfs; |
dce81290 | 2066 | |
d45f60c6 | 2067 | trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how); |
7f714720 | 2068 | if (trypnfs == PNFS_NOT_ATTEMPTED) |
d45f60c6 | 2069 | pnfs_write_through_mds(desc, hdr); |
dce81290 TM |
2070 | } |
2071 | ||
6c75dc0d FI |
2072 | static void pnfs_writehdr_free(struct nfs_pgio_header *hdr) |
2073 | { | |
9369a431 | 2074 | pnfs_put_lseg(hdr->lseg); |
1e7f3a48 | 2075 | nfs_pgio_header_free(hdr); |
6c75dc0d FI |
2076 | } |
2077 | ||
dce81290 TM |
2078 | int |
2079 | pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc) | |
2080 | { | |
6c75dc0d | 2081 | struct nfs_pgio_header *hdr; |
dce81290 TM |
2082 | int ret; |
2083 | ||
1e7f3a48 WAA |
2084 | hdr = nfs_pgio_header_alloc(desc->pg_rw_ops); |
2085 | if (!hdr) { | |
2bff2288 PT |
2086 | desc->pg_error = -ENOMEM; |
2087 | return desc->pg_error; | |
dce81290 | 2088 | } |
6c75dc0d | 2089 | nfs_pgheader_init(desc, hdr, pnfs_writehdr_free); |
180bb5ec | 2090 | |
9369a431 | 2091 | hdr->lseg = pnfs_get_lseg(desc->pg_lseg); |
ef2c488c | 2092 | ret = nfs_generic_pgio(desc, hdr); |
180bb5ec | 2093 | if (!ret) |
7f714720 | 2094 | pnfs_do_write(desc, hdr, desc->pg_ioflags); |
a7d42ddb | 2095 | |
6c75dc0d | 2096 | return ret; |
dce81290 TM |
2097 | } |
2098 | EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages); | |
2099 | ||
53113ad3 | 2100 | int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr) |
62e4a769 TM |
2101 | { |
2102 | struct nfs_pageio_descriptor pgio; | |
2103 | ||
1acbbb4e | 2104 | /* Resend all requests through the MDS */ |
53113ad3 WAA |
2105 | nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops); |
2106 | return nfs_pageio_resend(&pgio, hdr); | |
1acbbb4e | 2107 | } |
e7dd79af | 2108 | EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds); |
1acbbb4e | 2109 | |
d45f60c6 | 2110 | static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr) |
1acbbb4e | 2111 | { |
cd841605 FI |
2112 | dprintk("pnfs read error = %d\n", hdr->pnfs_error); |
2113 | if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags & | |
1acbbb4e | 2114 | PNFS_LAYOUTRET_ON_ERROR) { |
cd841605 | 2115 | pnfs_return_layout(hdr->inode); |
1acbbb4e | 2116 | } |
4db6e0b7 | 2117 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) |
53113ad3 | 2118 | hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr); |
62e4a769 TM |
2119 | } |
2120 | ||
d20581aa BH |
2121 | /* |
2122 | * Called by non rpc-based layout drivers | |
2123 | */ | |
d45f60c6 | 2124 | void pnfs_ld_read_done(struct nfs_pgio_header *hdr) |
d20581aa | 2125 | { |
cd841605 | 2126 | if (likely(!hdr->pnfs_error)) { |
d45f60c6 WAA |
2127 | __nfs4_read_done_cb(hdr); |
2128 | hdr->mds_ops->rpc_call_done(&hdr->task, hdr); | |
f8417b48 KM |
2129 | } |
2130 | trace_nfs4_pnfs_read(hdr, hdr->pnfs_error); | |
2131 | if (unlikely(hdr->pnfs_error)) | |
d45f60c6 WAA |
2132 | pnfs_ld_handle_read_error(hdr); |
2133 | hdr->mds_ops->rpc_release(hdr); | |
d20581aa BH |
2134 | } |
2135 | EXPORT_SYMBOL_GPL(pnfs_ld_read_done); | |
2136 | ||
493292dd TM |
2137 | static void |
2138 | pnfs_read_through_mds(struct nfs_pageio_descriptor *desc, | |
d45f60c6 | 2139 | struct nfs_pgio_header *hdr) |
493292dd | 2140 | { |
48d635f1 | 2141 | struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc); |
a7d42ddb | 2142 | |
4db6e0b7 | 2143 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { |
a7d42ddb | 2144 | list_splice_tail_init(&hdr->pages, &mirror->pg_list); |
4db6e0b7 | 2145 | nfs_pageio_reset_read_mds(desc); |
a7d42ddb | 2146 | mirror->pg_recoalesce = 1; |
4db6e0b7 | 2147 | } |
d45f60c6 | 2148 | nfs_pgio_data_destroy(hdr); |
1ca018d2 | 2149 | hdr->release(hdr); |
493292dd TM |
2150 | } |
2151 | ||
64419a9b AA |
2152 | /* |
2153 | * Call the appropriate parallel I/O subsystem read function. | |
2154 | */ | |
493292dd | 2155 | static enum pnfs_try_status |
d45f60c6 | 2156 | pnfs_try_to_read_data(struct nfs_pgio_header *hdr, |
493292dd TM |
2157 | const struct rpc_call_ops *call_ops, |
2158 | struct pnfs_layout_segment *lseg) | |
64419a9b | 2159 | { |
cd841605 | 2160 | struct inode *inode = hdr->inode; |
64419a9b AA |
2161 | struct nfs_server *nfss = NFS_SERVER(inode); |
2162 | enum pnfs_try_status trypnfs; | |
2163 | ||
cd841605 | 2164 | hdr->mds_ops = call_ops; |
64419a9b AA |
2165 | |
2166 | dprintk("%s: Reading ino:%lu %u@%llu\n", | |
d45f60c6 | 2167 | __func__, inode->i_ino, hdr->args.count, hdr->args.offset); |
64419a9b | 2168 | |
d45f60c6 | 2169 | trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr); |
4db6e0b7 | 2170 | if (trypnfs != PNFS_NOT_ATTEMPTED) |
64419a9b | 2171 | nfs_inc_stats(inode, NFSIOS_PNFS_READ); |
64419a9b AA |
2172 | dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs); |
2173 | return trypnfs; | |
2174 | } | |
863a3c6c | 2175 | |
ceb11e13 | 2176 | /* Resend all requests through pnfs. */ |
1b1bc66b | 2177 | void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr) |
ceb11e13 PT |
2178 | { |
2179 | struct nfs_pageio_descriptor pgio; | |
2180 | ||
1b1bc66b WAA |
2181 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { |
2182 | nfs_pageio_init_read(&pgio, hdr->inode, false, | |
2183 | hdr->completion_ops); | |
2184 | hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr); | |
2185 | } | |
ceb11e13 PT |
2186 | } |
2187 | EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs); | |
2188 | ||
493292dd | 2189 | static void |
7f714720 | 2190 | pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr) |
493292dd | 2191 | { |
493292dd TM |
2192 | const struct rpc_call_ops *call_ops = desc->pg_rpc_callops; |
2193 | struct pnfs_layout_segment *lseg = desc->pg_lseg; | |
7f714720 | 2194 | enum pnfs_try_status trypnfs; |
493292dd | 2195 | |
d45f60c6 | 2196 | trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg); |
ceb11e13 | 2197 | if (trypnfs == PNFS_TRY_AGAIN) |
1b1bc66b WAA |
2198 | pnfs_read_resend_pnfs(hdr); |
2199 | if (trypnfs == PNFS_NOT_ATTEMPTED || hdr->task.tk_status) | |
d45f60c6 | 2200 | pnfs_read_through_mds(desc, hdr); |
493292dd TM |
2201 | } |
2202 | ||
4db6e0b7 FI |
2203 | static void pnfs_readhdr_free(struct nfs_pgio_header *hdr) |
2204 | { | |
9369a431 | 2205 | pnfs_put_lseg(hdr->lseg); |
1e7f3a48 | 2206 | nfs_pgio_header_free(hdr); |
4db6e0b7 FI |
2207 | } |
2208 | ||
493292dd TM |
2209 | int |
2210 | pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc) | |
2211 | { | |
4db6e0b7 | 2212 | struct nfs_pgio_header *hdr; |
493292dd TM |
2213 | int ret; |
2214 | ||
1e7f3a48 WAA |
2215 | hdr = nfs_pgio_header_alloc(desc->pg_rw_ops); |
2216 | if (!hdr) { | |
2bff2288 PT |
2217 | desc->pg_error = -ENOMEM; |
2218 | return desc->pg_error; | |
493292dd | 2219 | } |
4db6e0b7 | 2220 | nfs_pgheader_init(desc, hdr, pnfs_readhdr_free); |
9369a431 | 2221 | hdr->lseg = pnfs_get_lseg(desc->pg_lseg); |
ef2c488c | 2222 | ret = nfs_generic_pgio(desc, hdr); |
180bb5ec | 2223 | if (!ret) |
7f714720 | 2224 | pnfs_do_read(desc, hdr); |
4db6e0b7 | 2225 | return ret; |
493292dd TM |
2226 | } |
2227 | EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages); | |
2228 | ||
71244d9b TM |
2229 | static void pnfs_clear_layoutcommitting(struct inode *inode) |
2230 | { | |
2231 | unsigned long *bitlock = &NFS_I(inode)->flags; | |
2232 | ||
2233 | clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock); | |
4e857c58 | 2234 | smp_mb__after_atomic(); |
71244d9b TM |
2235 | wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING); |
2236 | } | |
2237 | ||
863a3c6c | 2238 | /* |
a9bae566 | 2239 | * There can be multiple RW segments. |
863a3c6c | 2240 | */ |
a9bae566 | 2241 | static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp) |
863a3c6c | 2242 | { |
a9bae566 | 2243 | struct pnfs_layout_segment *lseg; |
863a3c6c | 2244 | |
a9bae566 PT |
2245 | list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) { |
2246 | if (lseg->pls_range.iomode == IOMODE_RW && | |
a073dbff | 2247 | test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) |
a9bae566 PT |
2248 | list_add(&lseg->pls_lc_list, listp); |
2249 | } | |
863a3c6c AA |
2250 | } |
2251 | ||
a073dbff TM |
2252 | static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp) |
2253 | { | |
2254 | struct pnfs_layout_segment *lseg, *tmp; | |
a073dbff TM |
2255 | |
2256 | /* Matched by references in pnfs_set_layoutcommit */ | |
2257 | list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) { | |
2258 | list_del_init(&lseg->pls_lc_list); | |
2259 | pnfs_put_lseg(lseg); | |
2260 | } | |
2261 | ||
71244d9b | 2262 | pnfs_clear_layoutcommitting(inode); |
a073dbff TM |
2263 | } |
2264 | ||
1b0ae068 PT |
2265 | void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg) |
2266 | { | |
b9e028fd | 2267 | pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode); |
1b0ae068 PT |
2268 | } |
2269 | EXPORT_SYMBOL_GPL(pnfs_set_lo_fail); | |
2270 | ||
863a3c6c | 2271 | void |
67af7611 TM |
2272 | pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg, |
2273 | loff_t end_pos) | |
863a3c6c | 2274 | { |
cd841605 | 2275 | struct nfs_inode *nfsi = NFS_I(inode); |
79a48a1f | 2276 | bool mark_as_dirty = false; |
863a3c6c | 2277 | |
cd841605 | 2278 | spin_lock(&inode->i_lock); |
863a3c6c | 2279 | if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) { |
29559b11 | 2280 | nfsi->layout->plh_lwb = end_pos; |
79a48a1f | 2281 | mark_as_dirty = true; |
863a3c6c | 2282 | dprintk("%s: Set layoutcommit for inode %lu ", |
cd841605 | 2283 | __func__, inode->i_ino); |
29559b11 TM |
2284 | } else if (end_pos > nfsi->layout->plh_lwb) |
2285 | nfsi->layout->plh_lwb = end_pos; | |
67af7611 | 2286 | if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) { |
a9bae566 | 2287 | /* references matched in nfs4_layoutcommit_release */ |
67af7611 | 2288 | pnfs_get_lseg(lseg); |
a9bae566 | 2289 | } |
cd841605 | 2290 | spin_unlock(&inode->i_lock); |
acff5880 | 2291 | dprintk("%s: lseg %p end_pos %llu\n", |
67af7611 | 2292 | __func__, lseg, nfsi->layout->plh_lwb); |
79a48a1f WAA |
2293 | |
2294 | /* if pnfs_layoutcommit_inode() runs between inode locks, the next one | |
2295 | * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */ | |
2296 | if (mark_as_dirty) | |
cd841605 | 2297 | mark_inode_dirty_sync(inode); |
863a3c6c AA |
2298 | } |
2299 | EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit); | |
2300 | ||
db29c089 AA |
2301 | void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data) |
2302 | { | |
2303 | struct nfs_server *nfss = NFS_SERVER(data->args.inode); | |
2304 | ||
2305 | if (nfss->pnfs_curr_ld->cleanup_layoutcommit) | |
2306 | nfss->pnfs_curr_ld->cleanup_layoutcommit(data); | |
a073dbff | 2307 | pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list); |
db29c089 AA |
2308 | } |
2309 | ||
de4b15c7 AA |
2310 | /* |
2311 | * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and | |
2312 | * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough | |
2313 | * data to disk to allow the server to recover the data if it crashes. | |
2314 | * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag | |
2315 | * is off, and a COMMIT is sent to a data server, or | |
2316 | * if WRITEs to a data server return NFS_DATA_SYNC. | |
2317 | */ | |
863a3c6c | 2318 | int |
ef311537 | 2319 | pnfs_layoutcommit_inode(struct inode *inode, bool sync) |
863a3c6c | 2320 | { |
5f919c9f | 2321 | struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld; |
863a3c6c AA |
2322 | struct nfs4_layoutcommit_data *data; |
2323 | struct nfs_inode *nfsi = NFS_I(inode); | |
863a3c6c | 2324 | loff_t end_pos; |
71244d9b | 2325 | int status; |
863a3c6c | 2326 | |
71244d9b | 2327 | if (!pnfs_layoutcommit_outstanding(inode)) |
de4b15c7 AA |
2328 | return 0; |
2329 | ||
71244d9b | 2330 | dprintk("--> %s inode %lu\n", __func__, inode->i_ino); |
92407e75 | 2331 | |
71244d9b | 2332 | status = -EAGAIN; |
92407e75 | 2333 | if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) { |
71244d9b TM |
2334 | if (!sync) |
2335 | goto out; | |
74316201 | 2336 | status = wait_on_bit_lock_action(&nfsi->flags, |
71244d9b TM |
2337 | NFS_INO_LAYOUTCOMMITTING, |
2338 | nfs_wait_bit_killable, | |
2339 | TASK_KILLABLE); | |
92407e75 | 2340 | if (status) |
71244d9b | 2341 | goto out; |
92407e75 PT |
2342 | } |
2343 | ||
71244d9b TM |
2344 | status = -ENOMEM; |
2345 | /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */ | |
2346 | data = kzalloc(sizeof(*data), GFP_NOFS); | |
2347 | if (!data) | |
2348 | goto clear_layoutcommitting; | |
2349 | ||
2350 | status = 0; | |
de4b15c7 | 2351 | spin_lock(&inode->i_lock); |
71244d9b TM |
2352 | if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) |
2353 | goto out_unlock; | |
a9bae566 | 2354 | |
71244d9b | 2355 | INIT_LIST_HEAD(&data->lseg_list); |
a9bae566 | 2356 | pnfs_list_write_lseg(inode, &data->lseg_list); |
863a3c6c | 2357 | |
acff5880 | 2358 | end_pos = nfsi->layout->plh_lwb; |
863a3c6c | 2359 | |
f597c537 | 2360 | nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid); |
863a3c6c AA |
2361 | spin_unlock(&inode->i_lock); |
2362 | ||
2363 | data->args.inode = inode; | |
9fa40758 | 2364 | data->cred = get_rpccred(nfsi->layout->plh_lc_cred); |
863a3c6c AA |
2365 | nfs_fattr_init(&data->fattr); |
2366 | data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask; | |
2367 | data->res.fattr = &data->fattr; | |
2368 | data->args.lastbytewritten = end_pos - 1; | |
2369 | data->res.server = NFS_SERVER(inode); | |
2370 | ||
5f919c9f CH |
2371 | if (ld->prepare_layoutcommit) { |
2372 | status = ld->prepare_layoutcommit(&data->args); | |
2373 | if (status) { | |
3471648a | 2374 | put_rpccred(data->cred); |
5f919c9f | 2375 | spin_lock(&inode->i_lock); |
29559b11 TM |
2376 | set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags); |
2377 | if (end_pos > nfsi->layout->plh_lwb) | |
5f919c9f | 2378 | nfsi->layout->plh_lwb = end_pos; |
3471648a | 2379 | goto out_unlock; |
5f919c9f CH |
2380 | } |
2381 | } | |
2382 | ||
2383 | ||
863a3c6c AA |
2384 | status = nfs4_proc_layoutcommit(data, sync); |
2385 | out: | |
92407e75 PT |
2386 | if (status) |
2387 | mark_inode_dirty_sync(inode); | |
863a3c6c AA |
2388 | dprintk("<-- %s status %d\n", __func__, status); |
2389 | return status; | |
71244d9b TM |
2390 | out_unlock: |
2391 | spin_unlock(&inode->i_lock); | |
92407e75 | 2392 | kfree(data); |
71244d9b TM |
2393 | clear_layoutcommitting: |
2394 | pnfs_clear_layoutcommitting(inode); | |
92407e75 | 2395 | goto out; |
863a3c6c | 2396 | } |
72cff449 | 2397 | EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode); |
82be417a | 2398 | |
5bb89b47 TM |
2399 | int |
2400 | pnfs_generic_sync(struct inode *inode, bool datasync) | |
2401 | { | |
2402 | return pnfs_layoutcommit_inode(inode, true); | |
2403 | } | |
2404 | EXPORT_SYMBOL_GPL(pnfs_generic_sync); | |
2405 | ||
82be417a AA |
2406 | struct nfs4_threshold *pnfs_mdsthreshold_alloc(void) |
2407 | { | |
2408 | struct nfs4_threshold *thp; | |
2409 | ||
2410 | thp = kzalloc(sizeof(*thp), GFP_NOFS); | |
2411 | if (!thp) { | |
2412 | dprintk("%s mdsthreshold allocation failed\n", __func__); | |
2413 | return NULL; | |
2414 | } | |
2415 | return thp; | |
2416 | } | |
8733408d | 2417 | |
865a7ecb | 2418 | #if IS_ENABLED(CONFIG_NFS_V4_2) |
8733408d | 2419 | int |
c8ad8894 | 2420 | pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags) |
8733408d PT |
2421 | { |
2422 | struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld; | |
2423 | struct nfs_server *server = NFS_SERVER(inode); | |
1bfe3b25 | 2424 | struct nfs_inode *nfsi = NFS_I(inode); |
8733408d PT |
2425 | struct nfs42_layoutstat_data *data; |
2426 | struct pnfs_layout_hdr *hdr; | |
2427 | int status = 0; | |
2428 | ||
2429 | if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats) | |
2430 | goto out; | |
2431 | ||
6c5a0d89 TM |
2432 | if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS)) |
2433 | goto out; | |
2434 | ||
1bfe3b25 PT |
2435 | if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags)) |
2436 | goto out; | |
2437 | ||
8733408d PT |
2438 | spin_lock(&inode->i_lock); |
2439 | if (!NFS_I(inode)->layout) { | |
2440 | spin_unlock(&inode->i_lock); | |
f538d0ba | 2441 | goto out_clear_layoutstats; |
8733408d PT |
2442 | } |
2443 | hdr = NFS_I(inode)->layout; | |
2444 | pnfs_get_layout_hdr(hdr); | |
2445 | spin_unlock(&inode->i_lock); | |
2446 | ||
c8ad8894 | 2447 | data = kzalloc(sizeof(*data), gfp_flags); |
8733408d PT |
2448 | if (!data) { |
2449 | status = -ENOMEM; | |
2450 | goto out_put; | |
2451 | } | |
2452 | ||
2453 | data->args.fh = NFS_FH(inode); | |
2454 | data->args.inode = inode; | |
2455 | nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid); | |
2456 | status = ld->prepare_layoutstats(&data->args); | |
2457 | if (status) | |
2458 | goto out_free; | |
2459 | ||
2460 | status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data); | |
2461 | ||
2462 | out: | |
2463 | dprintk("%s returns %d\n", __func__, status); | |
2464 | return status; | |
2465 | ||
2466 | out_free: | |
2467 | kfree(data); | |
2468 | out_put: | |
2469 | pnfs_put_layout_hdr(hdr); | |
f538d0ba | 2470 | out_clear_layoutstats: |
1bfe3b25 PT |
2471 | smp_mb__before_atomic(); |
2472 | clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags); | |
2473 | smp_mb__after_atomic(); | |
8733408d PT |
2474 | goto out; |
2475 | } | |
2476 | EXPORT_SYMBOL_GPL(pnfs_report_layoutstat); | |
865a7ecb | 2477 | #endif |
bbf58bf3 TM |
2478 | |
2479 | unsigned int layoutstats_timer; | |
2480 | module_param(layoutstats_timer, uint, 0644); | |
2481 | EXPORT_SYMBOL_GPL(layoutstats_timer); |