]>
Commit | Line | Data |
---|---|---|
d7e09d03 PT |
1 | /* |
2 | * GPL HEADER START | |
3 | * | |
4 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License version 2 only, | |
8 | * as published by the Free Software Foundation. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, but | |
11 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
13 | * General Public License version 2 for more details (a copy is included | |
14 | * in the LICENSE file that accompanied this code). | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * version 2 along with this program; If not, see | |
18 | * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf | |
19 | * | |
20 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
21 | * CA 95054 USA or visit www.sun.com if you need additional information or | |
22 | * have any questions. | |
23 | * | |
24 | * GPL HEADER END | |
25 | */ | |
26 | /* | |
27 | * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
30 | * Copyright (c) 2011, 2012, Intel Corporation. | |
31 | */ | |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | * | |
36 | * Implementation of cl_io for VVP layer. | |
37 | * | |
38 | * Author: Nikita Danilov <nikita.danilov@sun.com> | |
39 | * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com> | |
40 | */ | |
41 | ||
42 | #define DEBUG_SUBSYSTEM S_LLITE | |
43 | ||
44 | ||
67a235f5 GKH |
45 | #include "../include/obd.h" |
46 | #include "../include/lustre_lite.h" | |
d7e09d03 PT |
47 | |
48 | #include "vvp_internal.h" | |
49 | ||
50 | static struct vvp_io *cl2vvp_io(const struct lu_env *env, | |
51 | const struct cl_io_slice *slice); | |
52 | ||
53 | /** | |
74c0da19 | 54 | * True, if \a io is a normal io, False for splice_{read,write} |
d7e09d03 PT |
55 | */ |
56 | int cl_is_normalio(const struct lu_env *env, const struct cl_io *io) | |
57 | { | |
58 | struct vvp_io *vio = vvp_env_io(env); | |
59 | ||
60 | LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); | |
61 | ||
62 | return vio->cui_io_subtype == IO_NORMAL; | |
63 | } | |
64 | ||
65 | /** | |
66 | * For swapping layout. The file's layout may have changed. | |
67 | * To avoid populating pages to a wrong stripe, we have to verify the | |
68 | * correctness of layout. It works because swapping layout processes | |
69 | * have to acquire group lock. | |
70 | */ | |
71 | static bool can_populate_pages(const struct lu_env *env, struct cl_io *io, | |
72 | struct inode *inode) | |
73 | { | |
74 | struct ll_inode_info *lli = ll_i2info(inode); | |
75 | struct ccc_io *cio = ccc_env_io(env); | |
76 | bool rc = true; | |
77 | ||
78 | switch (io->ci_type) { | |
79 | case CIT_READ: | |
80 | case CIT_WRITE: | |
81 | /* don't need lock here to check lli_layout_gen as we have held | |
82 | * extent lock and GROUP lock has to hold to swap layout */ | |
09aed8a5 | 83 | if (ll_layout_version_get(lli) != cio->cui_layout_gen) { |
d7e09d03 PT |
84 | io->ci_need_restart = 1; |
85 | /* this will return application a short read/write */ | |
86 | io->ci_continue = 0; | |
87 | rc = false; | |
88 | } | |
89 | case CIT_FAULT: | |
90 | /* fault is okay because we've already had a page. */ | |
91 | default: | |
92 | break; | |
93 | } | |
94 | ||
95 | return rc; | |
96 | } | |
97 | ||
98 | /***************************************************************************** | |
99 | * | |
100 | * io operations. | |
101 | * | |
102 | */ | |
103 | ||
104 | static int vvp_io_fault_iter_init(const struct lu_env *env, | |
105 | const struct cl_io_slice *ios) | |
106 | { | |
107 | struct vvp_io *vio = cl2vvp_io(env, ios); | |
108 | struct inode *inode = ccc_object_inode(ios->cis_obj); | |
109 | ||
110 | LASSERT(inode == | |
2a8a3597 | 111 | file_inode(cl2ccc_io(env, ios)->cui_fd->fd_file)); |
d7e09d03 PT |
112 | vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime); |
113 | return 0; | |
114 | } | |
115 | ||
116 | static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios) | |
117 | { | |
118 | struct cl_io *io = ios->cis_io; | |
119 | struct cl_object *obj = io->ci_obj; | |
120 | struct ccc_io *cio = cl2ccc_io(env, ios); | |
121 | ||
122 | CLOBINVRNT(env, obj, ccc_object_invariant(obj)); | |
123 | ||
5ea17d6c JL |
124 | CDEBUG(D_VFSTRACE, DFID |
125 | " ignore/verify layout %d/%d, layout version %d restore needed %d\n", | |
126 | PFID(lu_object_fid(&obj->co_lu)), | |
127 | io->ci_ignore_layout, io->ci_verify_layout, | |
128 | cio->cui_layout_gen, io->ci_restore_needed); | |
129 | ||
130 | if (io->ci_restore_needed == 1) { | |
131 | int rc; | |
132 | ||
133 | /* file was detected release, we need to restore it | |
134 | * before finishing the io | |
135 | */ | |
136 | rc = ll_layout_restore(ccc_object_inode(obj)); | |
137 | /* if restore registration failed, no restart, | |
138 | * we will return -ENODATA */ | |
139 | /* The layout will change after restore, so we need to | |
140 | * block on layout lock hold by the MDT | |
141 | * as MDT will not send new layout in lvb (see LU-3124) | |
142 | * we have to explicitly fetch it, all this will be done | |
143 | * by ll_layout_refresh() | |
144 | */ | |
145 | if (rc == 0) { | |
146 | io->ci_restore_needed = 0; | |
147 | io->ci_need_restart = 1; | |
148 | io->ci_verify_layout = 1; | |
149 | } else { | |
150 | io->ci_restore_needed = 1; | |
151 | io->ci_need_restart = 0; | |
152 | io->ci_verify_layout = 0; | |
153 | io->ci_result = rc; | |
154 | } | |
155 | } | |
d7e09d03 PT |
156 | |
157 | if (!io->ci_ignore_layout && io->ci_verify_layout) { | |
158 | __u32 gen = 0; | |
159 | ||
160 | /* check layout version */ | |
161 | ll_layout_refresh(ccc_object_inode(obj), &gen); | |
162 | io->ci_need_restart = cio->cui_layout_gen != gen; | |
5ea17d6c JL |
163 | if (io->ci_need_restart) { |
164 | CDEBUG(D_VFSTRACE, | |
165 | DFID" layout changed from %d to %d.\n", | |
166 | PFID(lu_object_fid(&obj->co_lu)), | |
167 | cio->cui_layout_gen, gen); | |
168 | /* today successful restore is the only possible | |
169 | * case */ | |
170 | /* restore was done, clear restoring state */ | |
171 | ll_i2info(ccc_object_inode(obj))->lli_flags &= | |
172 | ~LLIF_FILE_RESTORING; | |
173 | } | |
d7e09d03 PT |
174 | } |
175 | } | |
176 | ||
177 | static void vvp_io_fault_fini(const struct lu_env *env, | |
178 | const struct cl_io_slice *ios) | |
179 | { | |
180 | struct cl_io *io = ios->cis_io; | |
181 | struct cl_page *page = io->u.ci_fault.ft_page; | |
182 | ||
183 | CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj)); | |
184 | ||
185 | if (page != NULL) { | |
186 | lu_ref_del(&page->cp_reference, "fault", io); | |
187 | cl_page_put(env, page); | |
188 | io->u.ci_fault.ft_page = NULL; | |
189 | } | |
190 | vvp_io_fini(env, ios); | |
191 | } | |
192 | ||
2d95f10e | 193 | static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma) |
d7e09d03 PT |
194 | { |
195 | /* | |
196 | * we only want to hold PW locks if the mmap() can generate | |
197 | * writes back to the file and that only happens in shared | |
198 | * writable vmas | |
199 | */ | |
200 | if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE)) | |
201 | return CLM_WRITE; | |
202 | return CLM_READ; | |
203 | } | |
204 | ||
205 | static int vvp_mmap_locks(const struct lu_env *env, | |
206 | struct ccc_io *vio, struct cl_io *io) | |
207 | { | |
208 | struct ccc_thread_info *cti = ccc_env_info(env); | |
209 | struct mm_struct *mm = current->mm; | |
210 | struct vm_area_struct *vma; | |
211 | struct cl_lock_descr *descr = &cti->cti_descr; | |
212 | ldlm_policy_data_t policy; | |
213 | unsigned long addr; | |
d7e09d03 PT |
214 | ssize_t count; |
215 | int result; | |
b42b15fd AV |
216 | struct iov_iter i; |
217 | struct iovec iov; | |
d7e09d03 PT |
218 | |
219 | LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); | |
220 | ||
221 | if (!cl_is_normalio(env, io)) | |
0a3bdb00 | 222 | return 0; |
d7e09d03 | 223 | |
b42b15fd | 224 | if (vio->cui_iter == NULL) /* nfs or loop back device write */ |
0a3bdb00 | 225 | return 0; |
d7e09d03 PT |
226 | |
227 | /* No MM (e.g. NFS)? No vmas too. */ | |
228 | if (mm == NULL) | |
0a3bdb00 | 229 | return 0; |
d7e09d03 | 230 | |
b42b15fd AV |
231 | iov_for_each(iov, i, *(vio->cui_iter)) { |
232 | addr = (unsigned long)iov.iov_base; | |
233 | count = iov.iov_len; | |
d7e09d03 PT |
234 | if (count == 0) |
235 | continue; | |
236 | ||
237 | count += addr & (~CFS_PAGE_MASK); | |
238 | addr &= CFS_PAGE_MASK; | |
239 | ||
240 | down_read(&mm->mmap_sem); | |
a58a38ac | 241 | while ((vma = our_vma(mm, addr, count)) != NULL) { |
2a8a3597 | 242 | struct inode *inode = file_inode(vma->vm_file); |
d7e09d03 PT |
243 | int flags = CEF_MUST; |
244 | ||
245 | if (ll_file_nolock(vma->vm_file)) { | |
246 | /* | |
247 | * For no lock case, a lockless lock will be | |
248 | * generated. | |
249 | */ | |
250 | flags = CEF_NEVER; | |
251 | } | |
252 | ||
253 | /* | |
254 | * XXX: Required lock mode can be weakened: CIT_WRITE | |
255 | * io only ever reads user level buffer, and CIT_READ | |
256 | * only writes on it. | |
257 | */ | |
258 | policy_from_vma(&policy, vma, addr, count); | |
259 | descr->cld_mode = vvp_mode_from_vma(vma); | |
260 | descr->cld_obj = ll_i2info(inode)->lli_clob; | |
261 | descr->cld_start = cl_index(descr->cld_obj, | |
262 | policy.l_extent.start); | |
263 | descr->cld_end = cl_index(descr->cld_obj, | |
264 | policy.l_extent.end); | |
265 | descr->cld_enq_flags = flags; | |
266 | result = cl_io_lock_alloc_add(env, io, descr); | |
267 | ||
268 | CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n", | |
269 | descr->cld_mode, descr->cld_start, | |
270 | descr->cld_end); | |
271 | ||
09516500 PF |
272 | if (result < 0) { |
273 | up_read(&mm->mmap_sem); | |
0a3bdb00 | 274 | return result; |
09516500 | 275 | } |
d7e09d03 PT |
276 | |
277 | if (vma->vm_end - addr >= count) | |
278 | break; | |
279 | ||
280 | count -= vma->vm_end - addr; | |
281 | addr = vma->vm_end; | |
282 | } | |
283 | up_read(&mm->mmap_sem); | |
284 | } | |
0a3bdb00 | 285 | return 0; |
d7e09d03 PT |
286 | } |
287 | ||
288 | static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io, | |
289 | enum cl_lock_mode mode, loff_t start, loff_t end) | |
290 | { | |
291 | struct ccc_io *cio = ccc_env_io(env); | |
292 | int result; | |
293 | int ast_flags = 0; | |
294 | ||
295 | LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); | |
d7e09d03 PT |
296 | |
297 | ccc_io_update_iov(env, cio, io); | |
298 | ||
299 | if (io->u.ci_rw.crw_nonblock) | |
300 | ast_flags |= CEF_NONBLOCK; | |
301 | result = vvp_mmap_locks(env, cio, io); | |
302 | if (result == 0) | |
303 | result = ccc_io_one_lock(env, io, ast_flags, mode, start, end); | |
0a3bdb00 | 304 | return result; |
d7e09d03 PT |
305 | } |
306 | ||
307 | static int vvp_io_read_lock(const struct lu_env *env, | |
308 | const struct cl_io_slice *ios) | |
309 | { | |
4c309612 JX |
310 | struct cl_io *io = ios->cis_io; |
311 | struct cl_io_rw_common *rd = &io->u.ci_rd.rd; | |
d7e09d03 PT |
312 | int result; |
313 | ||
4c309612 JX |
314 | result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos, |
315 | rd->crw_pos + rd->crw_count - 1); | |
316 | ||
0a3bdb00 | 317 | return result; |
d7e09d03 PT |
318 | } |
319 | ||
320 | static int vvp_io_fault_lock(const struct lu_env *env, | |
321 | const struct cl_io_slice *ios) | |
322 | { | |
323 | struct cl_io *io = ios->cis_io; | |
324 | struct vvp_io *vio = cl2vvp_io(env, ios); | |
325 | /* | |
326 | * XXX LDLM_FL_CBPENDING | |
327 | */ | |
328 | return ccc_io_one_lock_index | |
329 | (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma), | |
330 | io->u.ci_fault.ft_index, io->u.ci_fault.ft_index); | |
331 | } | |
332 | ||
333 | static int vvp_io_write_lock(const struct lu_env *env, | |
334 | const struct cl_io_slice *ios) | |
335 | { | |
336 | struct cl_io *io = ios->cis_io; | |
337 | loff_t start; | |
338 | loff_t end; | |
339 | ||
340 | if (io->u.ci_wr.wr_append) { | |
341 | start = 0; | |
342 | end = OBD_OBJECT_EOF; | |
343 | } else { | |
344 | start = io->u.ci_wr.wr.crw_pos; | |
345 | end = start + io->u.ci_wr.wr.crw_count - 1; | |
346 | } | |
347 | return vvp_io_rw_lock(env, io, CLM_WRITE, start, end); | |
348 | } | |
349 | ||
350 | static int vvp_io_setattr_iter_init(const struct lu_env *env, | |
351 | const struct cl_io_slice *ios) | |
352 | { | |
353 | return 0; | |
354 | } | |
355 | ||
356 | /** | |
357 | * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io. | |
358 | * | |
359 | * Handles "lockless io" mode when extent locking is done by server. | |
360 | */ | |
361 | static int vvp_io_setattr_lock(const struct lu_env *env, | |
362 | const struct cl_io_slice *ios) | |
363 | { | |
364 | struct ccc_io *cio = ccc_env_io(env); | |
365 | struct cl_io *io = ios->cis_io; | |
366 | __u64 new_size; | |
367 | __u32 enqflags = 0; | |
368 | ||
369 | if (cl_io_is_trunc(io)) { | |
370 | new_size = io->u.ci_setattr.sa_attr.lvb_size; | |
371 | if (new_size == 0) | |
372 | enqflags = CEF_DISCARD_DATA; | |
373 | } else { | |
374 | if ((io->u.ci_setattr.sa_attr.lvb_mtime >= | |
375 | io->u.ci_setattr.sa_attr.lvb_ctime) || | |
376 | (io->u.ci_setattr.sa_attr.lvb_atime >= | |
377 | io->u.ci_setattr.sa_attr.lvb_ctime)) | |
378 | return 0; | |
379 | new_size = 0; | |
380 | } | |
381 | cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK; | |
382 | return ccc_io_one_lock(env, io, enqflags, CLM_WRITE, | |
383 | new_size, OBD_OBJECT_EOF); | |
384 | } | |
385 | ||
386 | static int vvp_do_vmtruncate(struct inode *inode, size_t size) | |
387 | { | |
388 | int result; | |
389 | /* | |
390 | * Only ll_inode_size_lock is taken at this level. | |
391 | */ | |
392 | ll_inode_size_lock(inode); | |
393 | result = inode_newsize_ok(inode, size); | |
394 | if (result < 0) { | |
395 | ll_inode_size_unlock(inode); | |
396 | return result; | |
397 | } | |
398 | truncate_setsize(inode, size); | |
399 | ll_inode_size_unlock(inode); | |
400 | return result; | |
401 | } | |
402 | ||
403 | static int vvp_io_setattr_trunc(const struct lu_env *env, | |
404 | const struct cl_io_slice *ios, | |
405 | struct inode *inode, loff_t size) | |
406 | { | |
407 | inode_dio_wait(inode); | |
408 | return 0; | |
409 | } | |
410 | ||
411 | static int vvp_io_setattr_time(const struct lu_env *env, | |
412 | const struct cl_io_slice *ios) | |
413 | { | |
414 | struct cl_io *io = ios->cis_io; | |
415 | struct cl_object *obj = io->ci_obj; | |
416 | struct cl_attr *attr = ccc_env_thread_attr(env); | |
417 | int result; | |
418 | unsigned valid = CAT_CTIME; | |
419 | ||
420 | cl_object_attr_lock(obj); | |
421 | attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime; | |
422 | if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) { | |
423 | attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime; | |
424 | valid |= CAT_ATIME; | |
425 | } | |
426 | if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) { | |
427 | attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime; | |
428 | valid |= CAT_MTIME; | |
429 | } | |
430 | result = cl_object_attr_set(env, obj, attr, valid); | |
431 | cl_object_attr_unlock(obj); | |
432 | ||
433 | return result; | |
434 | } | |
435 | ||
436 | static int vvp_io_setattr_start(const struct lu_env *env, | |
437 | const struct cl_io_slice *ios) | |
438 | { | |
439 | struct cl_io *io = ios->cis_io; | |
440 | struct inode *inode = ccc_object_inode(io->ci_obj); | |
5dd16419 | 441 | int result = 0; |
d7e09d03 PT |
442 | |
443 | mutex_lock(&inode->i_mutex); | |
444 | if (cl_io_is_trunc(io)) | |
5dd16419 JX |
445 | result = vvp_io_setattr_trunc(env, ios, inode, |
446 | io->u.ci_setattr.sa_attr.lvb_size); | |
447 | if (result == 0) | |
448 | result = vvp_io_setattr_time(env, ios); | |
449 | return result; | |
d7e09d03 PT |
450 | } |
451 | ||
452 | static void vvp_io_setattr_end(const struct lu_env *env, | |
453 | const struct cl_io_slice *ios) | |
454 | { | |
455 | struct cl_io *io = ios->cis_io; | |
456 | struct inode *inode = ccc_object_inode(io->ci_obj); | |
457 | ||
81e053c7 | 458 | if (cl_io_is_trunc(io)) |
d7e09d03 PT |
459 | /* Truncate in memory pages - they must be clean pages |
460 | * because osc has already notified to destroy osc_extents. */ | |
461 | vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size); | |
81e053c7 | 462 | |
d7e09d03 PT |
463 | mutex_unlock(&inode->i_mutex); |
464 | } | |
465 | ||
466 | static void vvp_io_setattr_fini(const struct lu_env *env, | |
467 | const struct cl_io_slice *ios) | |
468 | { | |
469 | vvp_io_fini(env, ios); | |
470 | } | |
471 | ||
d7e09d03 PT |
472 | static int vvp_io_read_start(const struct lu_env *env, |
473 | const struct cl_io_slice *ios) | |
474 | { | |
475 | struct vvp_io *vio = cl2vvp_io(env, ios); | |
476 | struct ccc_io *cio = cl2ccc_io(env, ios); | |
477 | struct cl_io *io = ios->cis_io; | |
478 | struct cl_object *obj = io->ci_obj; | |
479 | struct inode *inode = ccc_object_inode(obj); | |
480 | struct ll_ra_read *bead = &vio->cui_bead; | |
481 | struct file *file = cio->cui_fd->fd_file; | |
482 | ||
483 | int result; | |
484 | loff_t pos = io->u.ci_rd.rd.crw_pos; | |
485 | long cnt = io->u.ci_rd.rd.crw_count; | |
486 | long tot = cio->cui_tot_count; | |
487 | int exceed = 0; | |
488 | ||
489 | CLOBINVRNT(env, obj, ccc_object_invariant(obj)); | |
490 | ||
491 | CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt); | |
492 | ||
493 | if (!can_populate_pages(env, io, inode)) | |
494 | return 0; | |
495 | ||
496 | result = ccc_prep_size(env, obj, io, pos, tot, &exceed); | |
497 | if (result != 0) | |
498 | return result; | |
499 | else if (exceed != 0) | |
500 | goto out; | |
501 | ||
502 | LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, | |
503 | "Read ino %lu, %lu bytes, offset %lld, size %llu\n", | |
504 | inode->i_ino, cnt, pos, i_size_read(inode)); | |
505 | ||
506 | /* turn off the kernel's read-ahead */ | |
507 | cio->cui_fd->fd_file->f_ra.ra_pages = 0; | |
508 | ||
509 | /* initialize read-ahead window once per syscall */ | |
510 | if (!vio->cui_ra_window_set) { | |
511 | vio->cui_ra_window_set = 1; | |
512 | bead->lrr_start = cl_index(obj, pos); | |
513 | /* | |
514 | * XXX: explicit PAGE_CACHE_SIZE | |
515 | */ | |
516 | bead->lrr_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1); | |
517 | ll_ra_read_in(file, bead); | |
518 | } | |
519 | ||
520 | /* BUG: 5972 */ | |
521 | file_accessed(file); | |
522 | switch (vio->cui_io_subtype) { | |
523 | case IO_NORMAL: | |
74c0da19 | 524 | LASSERT(cio->cui_iocb->ki_pos == pos); |
b42b15fd | 525 | result = generic_file_read_iter(cio->cui_iocb, cio->cui_iter); |
74c0da19 | 526 | break; |
d7e09d03 PT |
527 | case IO_SPLICE: |
528 | result = generic_file_splice_read(file, &pos, | |
529 | vio->u.splice.cui_pipe, cnt, | |
530 | vio->u.splice.cui_flags); | |
531 | /* LU-1109: do splice read stripe by stripe otherwise if it | |
532 | * may make nfsd stuck if this read occupied all internal pipe | |
533 | * buffers. */ | |
534 | io->ci_continue = 0; | |
535 | break; | |
536 | default: | |
537 | CERROR("Wrong IO type %u\n", vio->cui_io_subtype); | |
538 | LBUG(); | |
539 | } | |
540 | ||
541 | out: | |
542 | if (result >= 0) { | |
543 | if (result < cnt) | |
544 | io->ci_continue = 0; | |
545 | io->ci_nob += result; | |
546 | ll_rw_stats_tally(ll_i2sbi(inode), current->pid, | |
4f37bc04 | 547 | cio->cui_fd, pos, result, READ); |
d7e09d03 PT |
548 | result = 0; |
549 | } | |
550 | return result; | |
551 | } | |
552 | ||
553 | static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios) | |
554 | { | |
555 | struct vvp_io *vio = cl2vvp_io(env, ios); | |
556 | struct ccc_io *cio = cl2ccc_io(env, ios); | |
557 | ||
558 | if (vio->cui_ra_window_set) | |
559 | ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead); | |
560 | ||
561 | vvp_io_fini(env, ios); | |
562 | } | |
563 | ||
564 | static int vvp_io_write_start(const struct lu_env *env, | |
565 | const struct cl_io_slice *ios) | |
566 | { | |
567 | struct ccc_io *cio = cl2ccc_io(env, ios); | |
568 | struct cl_io *io = ios->cis_io; | |
569 | struct cl_object *obj = io->ci_obj; | |
570 | struct inode *inode = ccc_object_inode(obj); | |
d7e09d03 PT |
571 | ssize_t result = 0; |
572 | loff_t pos = io->u.ci_wr.wr.crw_pos; | |
573 | size_t cnt = io->u.ci_wr.wr.crw_count; | |
574 | ||
d7e09d03 PT |
575 | if (!can_populate_pages(env, io, inode)) |
576 | return 0; | |
577 | ||
578 | if (cl_io_is_append(io)) { | |
579 | /* | |
580 | * PARALLEL IO This has to be changed for parallel IO doing | |
581 | * out-of-order writes. | |
582 | */ | |
583 | pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode); | |
584 | cio->cui_iocb->ki_pos = pos; | |
74c0da19 JX |
585 | } else { |
586 | LASSERT(cio->cui_iocb->ki_pos == pos); | |
d7e09d03 PT |
587 | } |
588 | ||
589 | CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt); | |
590 | ||
b42b15fd | 591 | if (cio->cui_iter == NULL) /* from a temp io in ll_cl_init(). */ |
d7e09d03 PT |
592 | result = 0; |
593 | else | |
b42b15fd AV |
594 | result = generic_file_write_iter(cio->cui_iocb, cio->cui_iter); |
595 | ||
d7e09d03 PT |
596 | if (result > 0) { |
597 | if (result < cnt) | |
598 | io->ci_continue = 0; | |
599 | io->ci_nob += result; | |
600 | ll_rw_stats_tally(ll_i2sbi(inode), current->pid, | |
4f37bc04 | 601 | cio->cui_fd, pos, result, WRITE); |
d7e09d03 PT |
602 | result = 0; |
603 | } | |
0a3bdb00 | 604 | return result; |
d7e09d03 PT |
605 | } |
606 | ||
607 | static int vvp_io_kernel_fault(struct vvp_fault_io *cfio) | |
608 | { | |
609 | struct vm_fault *vmf = cfio->fault.ft_vmf; | |
610 | ||
611 | cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf); | |
6aa51072 | 612 | cfio->fault.ft_flags_valid = 1; |
d7e09d03 PT |
613 | |
614 | if (vmf->page) { | |
aa3bee0d GKH |
615 | CDEBUG(D_PAGE, |
616 | "page %p map %p index %lu flags %lx count %u priv %0lx: got addr %p type NOPAGE\n", | |
617 | vmf->page, vmf->page->mapping, vmf->page->index, | |
618 | (long)vmf->page->flags, page_count(vmf->page), | |
619 | page_private(vmf->page), vmf->virtual_address); | |
d7e09d03 PT |
620 | if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) { |
621 | lock_page(vmf->page); | |
733bd244 | 622 | cfio->fault.ft_flags |= VM_FAULT_LOCKED; |
d7e09d03 PT |
623 | } |
624 | ||
625 | cfio->ft_vmpage = vmf->page; | |
626 | return 0; | |
627 | } | |
628 | ||
33692f27 | 629 | if (cfio->fault.ft_flags & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV)) { |
d7e09d03 PT |
630 | CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address); |
631 | return -EFAULT; | |
632 | } | |
633 | ||
634 | if (cfio->fault.ft_flags & VM_FAULT_OOM) { | |
635 | CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address); | |
636 | return -ENOMEM; | |
637 | } | |
638 | ||
639 | if (cfio->fault.ft_flags & VM_FAULT_RETRY) | |
640 | return -EAGAIN; | |
641 | ||
d0a0acc3 | 642 | CERROR("Unknown error in page fault %d!\n", cfio->fault.ft_flags); |
d7e09d03 PT |
643 | return -EINVAL; |
644 | } | |
645 | ||
646 | ||
647 | static int vvp_io_fault_start(const struct lu_env *env, | |
648 | const struct cl_io_slice *ios) | |
649 | { | |
650 | struct vvp_io *vio = cl2vvp_io(env, ios); | |
651 | struct cl_io *io = ios->cis_io; | |
652 | struct cl_object *obj = io->ci_obj; | |
653 | struct inode *inode = ccc_object_inode(obj); | |
654 | struct cl_fault_io *fio = &io->u.ci_fault; | |
655 | struct vvp_fault_io *cfio = &vio->u.fault; | |
656 | loff_t offset; | |
657 | int result = 0; | |
658 | struct page *vmpage = NULL; | |
659 | struct cl_page *page; | |
660 | loff_t size; | |
661 | pgoff_t last; /* last page in a file data region */ | |
662 | ||
663 | if (fio->ft_executable && | |
664 | LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime) | |
665 | CWARN("binary "DFID | |
666 | " changed while waiting for the page fault lock\n", | |
667 | PFID(lu_object_fid(&obj->co_lu))); | |
668 | ||
669 | /* offset of the last byte on the page */ | |
670 | offset = cl_offset(obj, fio->ft_index + 1) - 1; | |
671 | LASSERT(cl_index(obj, offset) == fio->ft_index); | |
672 | result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL); | |
673 | if (result != 0) | |
674 | return result; | |
675 | ||
676 | /* must return locked page */ | |
677 | if (fio->ft_mkwrite) { | |
678 | LASSERT(cfio->ft_vmpage != NULL); | |
679 | lock_page(cfio->ft_vmpage); | |
680 | } else { | |
681 | result = vvp_io_kernel_fault(cfio); | |
682 | if (result != 0) | |
683 | return result; | |
684 | } | |
685 | ||
686 | vmpage = cfio->ft_vmpage; | |
687 | LASSERT(PageLocked(vmpage)); | |
688 | ||
689 | if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE)) | |
690 | ll_invalidate_page(vmpage); | |
691 | ||
692 | size = i_size_read(inode); | |
693 | /* Though we have already held a cl_lock upon this page, but | |
694 | * it still can be truncated locally. */ | |
695 | if (unlikely((vmpage->mapping != inode->i_mapping) || | |
696 | (page_offset(vmpage) > size))) { | |
697 | CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n"); | |
698 | ||
699 | /* return +1 to stop cl_io_loop() and ll_fault() will catch | |
700 | * and retry. */ | |
34e1f2bb JL |
701 | result = +1; |
702 | goto out; | |
d7e09d03 PT |
703 | } |
704 | ||
705 | ||
557732ad | 706 | if (fio->ft_mkwrite) { |
d7e09d03 PT |
707 | pgoff_t last_index; |
708 | /* | |
709 | * Capture the size while holding the lli_trunc_sem from above | |
710 | * we want to make sure that we complete the mkwrite action | |
711 | * while holding this lock. We need to make sure that we are | |
712 | * not past the end of the file. | |
713 | */ | |
714 | last_index = cl_index(obj, size - 1); | |
715 | if (last_index < fio->ft_index) { | |
716 | CDEBUG(D_PAGE, | |
2d00bd17 JP |
717 | "llite: mkwrite and truncate race happened: %p: 0x%lx 0x%lx\n", |
718 | vmpage->mapping, fio->ft_index, last_index); | |
d7e09d03 PT |
719 | /* |
720 | * We need to return if we are | |
721 | * passed the end of the file. This will propagate | |
722 | * up the call stack to ll_page_mkwrite where | |
723 | * we will return VM_FAULT_NOPAGE. Any non-negative | |
724 | * value returned here will be silently | |
725 | * converted to 0. If the vmpage->mapping is null | |
726 | * the error code would be converted back to ENODATA | |
727 | * in ll_page_mkwrite0. Thus we return -ENODATA | |
728 | * to handle both cases | |
729 | */ | |
34e1f2bb JL |
730 | result = -ENODATA; |
731 | goto out; | |
d7e09d03 PT |
732 | } |
733 | } | |
734 | ||
735 | page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE); | |
34e1f2bb JL |
736 | if (IS_ERR(page)) { |
737 | result = PTR_ERR(page); | |
738 | goto out; | |
739 | } | |
d7e09d03 PT |
740 | |
741 | /* if page is going to be written, we should add this page into cache | |
742 | * earlier. */ | |
743 | if (fio->ft_mkwrite) { | |
744 | wait_on_page_writeback(vmpage); | |
745 | if (set_page_dirty(vmpage)) { | |
746 | struct ccc_page *cp; | |
747 | ||
748 | /* vvp_page_assume() calls wait_on_page_writeback(). */ | |
749 | cl_page_assume(env, io, page); | |
750 | ||
751 | cp = cl2ccc_page(cl_page_at(page, &vvp_device_type)); | |
752 | vvp_write_pending(cl2ccc(obj), cp); | |
753 | ||
754 | /* Do not set Dirty bit here so that in case IO is | |
755 | * started before the page is really made dirty, we | |
756 | * still have chance to detect it. */ | |
757 | result = cl_page_cache_add(env, io, page, CRT_WRITE); | |
758 | LASSERT(cl_page_is_owned(page, io)); | |
759 | ||
760 | vmpage = NULL; | |
761 | if (result < 0) { | |
762 | cl_page_unmap(env, io, page); | |
763 | cl_page_discard(env, io, page); | |
764 | cl_page_disown(env, io, page); | |
765 | ||
766 | cl_page_put(env, page); | |
767 | ||
768 | /* we're in big trouble, what can we do now? */ | |
769 | if (result == -EDQUOT) | |
770 | result = -ENOSPC; | |
34e1f2bb | 771 | goto out; |
d7e09d03 PT |
772 | } else |
773 | cl_page_disown(env, io, page); | |
774 | } | |
775 | } | |
776 | ||
777 | last = cl_index(obj, size - 1); | |
778 | /* | |
779 | * The ft_index is only used in the case of | |
780 | * a mkwrite action. We need to check | |
781 | * our assertions are correct, since | |
782 | * we should have caught this above | |
783 | */ | |
784 | LASSERT(!fio->ft_mkwrite || fio->ft_index <= last); | |
785 | if (fio->ft_index == last) | |
786 | /* | |
787 | * Last page is mapped partially. | |
788 | */ | |
789 | fio->ft_nob = size - cl_offset(obj, fio->ft_index); | |
790 | else | |
791 | fio->ft_nob = cl_page_size(obj); | |
792 | ||
793 | lu_ref_add(&page->cp_reference, "fault", io); | |
794 | fio->ft_page = page; | |
d7e09d03 PT |
795 | |
796 | out: | |
797 | /* return unlocked vmpage to avoid deadlocking */ | |
798 | if (vmpage != NULL) | |
799 | unlock_page(vmpage); | |
800 | cfio->fault.ft_flags &= ~VM_FAULT_LOCKED; | |
801 | return result; | |
802 | } | |
803 | ||
804 | static int vvp_io_fsync_start(const struct lu_env *env, | |
805 | const struct cl_io_slice *ios) | |
806 | { | |
807 | /* we should mark TOWRITE bit to each dirty page in radix tree to | |
808 | * verify pages have been written, but this is difficult because of | |
809 | * race. */ | |
810 | return 0; | |
811 | } | |
812 | ||
813 | static int vvp_io_read_page(const struct lu_env *env, | |
814 | const struct cl_io_slice *ios, | |
815 | const struct cl_page_slice *slice) | |
816 | { | |
817 | struct cl_io *io = ios->cis_io; | |
818 | struct cl_object *obj = slice->cpl_obj; | |
819 | struct ccc_page *cp = cl2ccc_page(slice); | |
820 | struct cl_page *page = slice->cpl_page; | |
821 | struct inode *inode = ccc_object_inode(obj); | |
822 | struct ll_sb_info *sbi = ll_i2sbi(inode); | |
823 | struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd; | |
824 | struct ll_readahead_state *ras = &fd->fd_ras; | |
825 | struct page *vmpage = cp->cpg_page; | |
826 | struct cl_2queue *queue = &io->ci_queue; | |
827 | int rc; | |
828 | ||
829 | CLOBINVRNT(env, obj, ccc_object_invariant(obj)); | |
830 | LASSERT(slice->cpl_obj == obj); | |
831 | ||
d7e09d03 PT |
832 | if (sbi->ll_ra_info.ra_max_pages_per_file && |
833 | sbi->ll_ra_info.ra_max_pages) | |
834 | ras_update(sbi, inode, ras, page->cp_index, | |
835 | cp->cpg_defer_uptodate); | |
836 | ||
837 | /* Sanity check whether the page is protected by a lock. */ | |
838 | rc = cl_page_is_under_lock(env, io, page); | |
839 | if (rc != -EBUSY) { | |
840 | CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n", | |
841 | rc == -ENODATA ? "without a lock" : | |
842 | "match failed", rc); | |
843 | if (rc != -ENODATA) | |
0a3bdb00 | 844 | return rc; |
d7e09d03 PT |
845 | } |
846 | ||
847 | if (cp->cpg_defer_uptodate) { | |
848 | cp->cpg_ra_used = 1; | |
849 | cl_page_export(env, page, 1); | |
850 | } | |
851 | /* | |
852 | * Add page into the queue even when it is marked uptodate above. | |
853 | * this will unlock it automatically as part of cl_page_list_disown(). | |
854 | */ | |
855 | cl_2queue_add(queue, page); | |
856 | if (sbi->ll_ra_info.ra_max_pages_per_file && | |
857 | sbi->ll_ra_info.ra_max_pages) | |
858 | ll_readahead(env, io, ras, | |
859 | vmpage->mapping, &queue->c2_qin, fd->fd_flags); | |
860 | ||
0a3bdb00 | 861 | return 0; |
d7e09d03 PT |
862 | } |
863 | ||
864 | static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io, | |
865 | struct cl_page *page, struct ccc_page *cp, | |
866 | enum cl_req_type crt) | |
867 | { | |
868 | struct cl_2queue *queue; | |
869 | int result; | |
870 | ||
871 | LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); | |
872 | ||
873 | queue = &io->ci_queue; | |
874 | cl_2queue_init_page(queue, page); | |
875 | ||
876 | result = cl_io_submit_sync(env, io, crt, queue, 0); | |
877 | LASSERT(cl_page_is_owned(page, io)); | |
878 | ||
879 | if (crt == CRT_READ) | |
880 | /* | |
881 | * in CRT_WRITE case page is left locked even in case of | |
882 | * error. | |
883 | */ | |
884 | cl_page_list_disown(env, io, &queue->c2_qin); | |
885 | cl_2queue_fini(env, queue); | |
886 | ||
887 | return result; | |
888 | } | |
889 | ||
890 | /** | |
891 | * Prepare partially written-to page for a write. | |
892 | */ | |
893 | static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io, | |
894 | struct cl_object *obj, struct cl_page *pg, | |
895 | struct ccc_page *cp, | |
896 | unsigned from, unsigned to) | |
897 | { | |
898 | struct cl_attr *attr = ccc_env_thread_attr(env); | |
899 | loff_t offset = cl_offset(obj, pg->cp_index); | |
900 | int result; | |
901 | ||
902 | cl_object_attr_lock(obj); | |
903 | result = cl_object_attr_get(env, obj, attr); | |
904 | cl_object_attr_unlock(obj); | |
905 | if (result == 0) { | |
906 | /* | |
907 | * If are writing to a new page, no need to read old data. | |
908 | * The extent locking will have updated the KMS, and for our | |
909 | * purposes here we can treat it like i_size. | |
910 | */ | |
911 | if (attr->cat_kms <= offset) { | |
5e8ebf13 | 912 | char *kaddr = kmap_atomic(cp->cpg_page); |
d7e09d03 PT |
913 | |
914 | memset(kaddr, 0, cl_page_size(obj)); | |
5e8ebf13 | 915 | kunmap_atomic(kaddr); |
d7e09d03 PT |
916 | } else if (cp->cpg_defer_uptodate) |
917 | cp->cpg_ra_used = 1; | |
918 | else | |
919 | result = vvp_page_sync_io(env, io, pg, cp, CRT_READ); | |
920 | /* | |
921 | * In older implementations, obdo_refresh_inode is called here | |
922 | * to update the inode because the write might modify the | |
923 | * object info at OST. However, this has been proven useless, | |
924 | * since LVB functions will be called when user space program | |
925 | * tries to retrieve inode attribute. Also, see bug 15909 for | |
926 | * details. -jay | |
927 | */ | |
928 | if (result == 0) | |
929 | cl_page_export(env, pg, 1); | |
930 | } | |
931 | return result; | |
932 | } | |
933 | ||
934 | static int vvp_io_prepare_write(const struct lu_env *env, | |
935 | const struct cl_io_slice *ios, | |
936 | const struct cl_page_slice *slice, | |
937 | unsigned from, unsigned to) | |
938 | { | |
939 | struct cl_object *obj = slice->cpl_obj; | |
940 | struct ccc_page *cp = cl2ccc_page(slice); | |
941 | struct cl_page *pg = slice->cpl_page; | |
942 | struct page *vmpage = cp->cpg_page; | |
943 | ||
944 | int result; | |
945 | ||
d7e09d03 PT |
946 | LINVRNT(cl_page_is_vmlocked(env, pg)); |
947 | LASSERT(vmpage->mapping->host == ccc_object_inode(obj)); | |
948 | ||
949 | result = 0; | |
950 | ||
951 | CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to); | |
952 | if (!PageUptodate(vmpage)) { | |
953 | /* | |
954 | * We're completely overwriting an existing page, so _don't_ | |
955 | * set it up to date until commit_write | |
956 | */ | |
957 | if (from == 0 && to == PAGE_CACHE_SIZE) { | |
958 | CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n"); | |
959 | POISON_PAGE(page, 0x11); | |
960 | } else | |
961 | result = vvp_io_prepare_partial(env, ios->cis_io, obj, | |
962 | pg, cp, from, to); | |
963 | } else | |
964 | CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n"); | |
0a3bdb00 | 965 | return result; |
d7e09d03 PT |
966 | } |
967 | ||
968 | static int vvp_io_commit_write(const struct lu_env *env, | |
969 | const struct cl_io_slice *ios, | |
970 | const struct cl_page_slice *slice, | |
971 | unsigned from, unsigned to) | |
972 | { | |
973 | struct cl_object *obj = slice->cpl_obj; | |
974 | struct cl_io *io = ios->cis_io; | |
975 | struct ccc_page *cp = cl2ccc_page(slice); | |
976 | struct cl_page *pg = slice->cpl_page; | |
977 | struct inode *inode = ccc_object_inode(obj); | |
978 | struct ll_sb_info *sbi = ll_i2sbi(inode); | |
979 | struct ll_inode_info *lli = ll_i2info(inode); | |
980 | struct page *vmpage = cp->cpg_page; | |
981 | ||
982 | int result; | |
983 | int tallyop; | |
984 | loff_t size; | |
985 | ||
d7e09d03 PT |
986 | LINVRNT(cl_page_is_vmlocked(env, pg)); |
987 | LASSERT(vmpage->mapping->host == inode); | |
988 | ||
c4f39553 | 989 | LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "committing page write\n"); |
d7e09d03 PT |
990 | CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to); |
991 | ||
992 | /* | |
993 | * queue a write for some time in the future the first time we | |
994 | * dirty the page. | |
995 | * | |
996 | * This is different from what other file systems do: they usually | |
997 | * just mark page (and some of its buffers) dirty and rely on | |
998 | * balance_dirty_pages() to start a write-back. Lustre wants write-back | |
999 | * to be started earlier for the following reasons: | |
1000 | * | |
1001 | * (1) with a large number of clients we need to limit the amount | |
1002 | * of cached data on the clients a lot; | |
1003 | * | |
1004 | * (2) large compute jobs generally want compute-only then io-only | |
1005 | * and the IO should complete as quickly as possible; | |
1006 | * | |
1007 | * (3) IO is batched up to the RPC size and is async until the | |
1008 | * client max cache is hit | |
1009 | * (/proc/fs/lustre/osc/OSC.../max_dirty_mb) | |
1010 | * | |
1011 | */ | |
1012 | if (!PageDirty(vmpage)) { | |
1013 | tallyop = LPROC_LL_DIRTY_MISSES; | |
1014 | result = cl_page_cache_add(env, io, pg, CRT_WRITE); | |
1015 | if (result == 0) { | |
1016 | /* page was added into cache successfully. */ | |
1017 | set_page_dirty(vmpage); | |
1018 | vvp_write_pending(cl2ccc(obj), cp); | |
1019 | } else if (result == -EDQUOT) { | |
1020 | pgoff_t last_index = i_size_read(inode) >> PAGE_CACHE_SHIFT; | |
1021 | bool need_clip = true; | |
1022 | ||
1023 | /* | |
1024 | * Client ran out of disk space grant. Possible | |
1025 | * strategies are: | |
1026 | * | |
1027 | * (a) do a sync write, renewing grant; | |
1028 | * | |
1029 | * (b) stop writing on this stripe, switch to the | |
1030 | * next one. | |
1031 | * | |
1032 | * (b) is a part of "parallel io" design that is the | |
1033 | * ultimate goal. (a) is what "old" client did, and | |
1034 | * what the new code continues to do for the time | |
1035 | * being. | |
1036 | */ | |
1037 | if (last_index > pg->cp_index) { | |
1038 | to = PAGE_CACHE_SIZE; | |
1039 | need_clip = false; | |
1040 | } else if (last_index == pg->cp_index) { | |
1041 | int size_to = i_size_read(inode) & ~CFS_PAGE_MASK; | |
1042 | if (to < size_to) | |
1043 | to = size_to; | |
1044 | } | |
1045 | if (need_clip) | |
1046 | cl_page_clip(env, pg, 0, to); | |
1047 | result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE); | |
1048 | if (result) | |
1049 | CERROR("Write page %lu of inode %p failed %d\n", | |
1050 | pg->cp_index, inode, result); | |
1051 | } | |
1052 | } else { | |
1053 | tallyop = LPROC_LL_DIRTY_HITS; | |
1054 | result = 0; | |
1055 | } | |
1056 | ll_stats_ops_tally(sbi, tallyop, 1); | |
1057 | ||
1058 | /* Inode should be marked DIRTY even if no new page was marked DIRTY | |
1059 | * because page could have been not flushed between 2 modifications. | |
1060 | * It is important the file is marked DIRTY as soon as the I/O is done | |
1061 | * Indeed, when cache is flushed, file could be already closed and it | |
1062 | * is too late to warn the MDT. | |
1063 | * It is acceptable that file is marked DIRTY even if I/O is dropped | |
1064 | * for some reasons before being flushed to OST. | |
1065 | */ | |
1066 | if (result == 0) { | |
1067 | spin_lock(&lli->lli_lock); | |
1068 | lli->lli_flags |= LLIF_DATA_MODIFIED; | |
1069 | spin_unlock(&lli->lli_lock); | |
1070 | } | |
1071 | ||
1072 | size = cl_offset(obj, pg->cp_index) + to; | |
1073 | ||
1074 | ll_inode_size_lock(inode); | |
1075 | if (result == 0) { | |
1076 | if (size > i_size_read(inode)) { | |
1077 | cl_isize_write_nolock(inode, size); | |
1078 | CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n", | |
1079 | PFID(lu_object_fid(&obj->co_lu)), | |
1080 | (unsigned long)size); | |
1081 | } | |
1082 | cl_page_export(env, pg, 1); | |
1083 | } else { | |
1084 | if (size > i_size_read(inode)) | |
1085 | cl_page_discard(env, io, pg); | |
1086 | } | |
1087 | ll_inode_size_unlock(inode); | |
0a3bdb00 | 1088 | return result; |
d7e09d03 PT |
1089 | } |
1090 | ||
1091 | static const struct cl_io_operations vvp_io_ops = { | |
1092 | .op = { | |
1093 | [CIT_READ] = { | |
1094 | .cio_fini = vvp_io_read_fini, | |
1095 | .cio_lock = vvp_io_read_lock, | |
1096 | .cio_start = vvp_io_read_start, | |
1097 | .cio_advance = ccc_io_advance | |
1098 | }, | |
1099 | [CIT_WRITE] = { | |
1100 | .cio_fini = vvp_io_fini, | |
1101 | .cio_lock = vvp_io_write_lock, | |
1102 | .cio_start = vvp_io_write_start, | |
1103 | .cio_advance = ccc_io_advance | |
1104 | }, | |
1105 | [CIT_SETATTR] = { | |
1106 | .cio_fini = vvp_io_setattr_fini, | |
1107 | .cio_iter_init = vvp_io_setattr_iter_init, | |
1108 | .cio_lock = vvp_io_setattr_lock, | |
1109 | .cio_start = vvp_io_setattr_start, | |
1110 | .cio_end = vvp_io_setattr_end | |
1111 | }, | |
1112 | [CIT_FAULT] = { | |
1113 | .cio_fini = vvp_io_fault_fini, | |
1114 | .cio_iter_init = vvp_io_fault_iter_init, | |
1115 | .cio_lock = vvp_io_fault_lock, | |
1116 | .cio_start = vvp_io_fault_start, | |
1117 | .cio_end = ccc_io_end | |
1118 | }, | |
1119 | [CIT_FSYNC] = { | |
1120 | .cio_start = vvp_io_fsync_start, | |
1121 | .cio_fini = vvp_io_fini | |
1122 | }, | |
1123 | [CIT_MISC] = { | |
1124 | .cio_fini = vvp_io_fini | |
1125 | } | |
1126 | }, | |
1127 | .cio_read_page = vvp_io_read_page, | |
1128 | .cio_prepare_write = vvp_io_prepare_write, | |
1129 | .cio_commit_write = vvp_io_commit_write | |
1130 | }; | |
1131 | ||
1132 | int vvp_io_init(const struct lu_env *env, struct cl_object *obj, | |
1133 | struct cl_io *io) | |
1134 | { | |
1135 | struct vvp_io *vio = vvp_env_io(env); | |
1136 | struct ccc_io *cio = ccc_env_io(env); | |
1137 | struct inode *inode = ccc_object_inode(obj); | |
1138 | int result; | |
1139 | ||
1140 | CLOBINVRNT(env, obj, ccc_object_invariant(obj)); | |
d7e09d03 | 1141 | |
5ea17d6c JL |
1142 | CDEBUG(D_VFSTRACE, DFID |
1143 | " ignore/verify layout %d/%d, layout version %d restore needed %d\n", | |
1144 | PFID(lu_object_fid(&obj->co_lu)), | |
1145 | io->ci_ignore_layout, io->ci_verify_layout, | |
1146 | cio->cui_layout_gen, io->ci_restore_needed); | |
1147 | ||
d7e09d03 PT |
1148 | CL_IO_SLICE_CLEAN(cio, cui_cl); |
1149 | cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops); | |
1150 | vio->cui_ra_window_set = 0; | |
1151 | result = 0; | |
1152 | if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) { | |
1153 | size_t count; | |
1154 | struct ll_inode_info *lli = ll_i2info(inode); | |
1155 | ||
1156 | count = io->u.ci_rw.crw_count; | |
1157 | /* "If nbyte is 0, read() will return 0 and have no other | |
1158 | * results." -- Single Unix Spec */ | |
1159 | if (count == 0) | |
1160 | result = 1; | |
b42b15fd | 1161 | else |
d7e09d03 | 1162 | cio->cui_tot_count = count; |
b42b15fd | 1163 | |
d7e09d03 PT |
1164 | /* for read/write, we store the jobid in the inode, and |
1165 | * it'll be fetched by osc when building RPC. | |
1166 | * | |
1167 | * it's not accurate if the file is shared by different | |
1168 | * jobs. | |
1169 | */ | |
1170 | lustre_get_jobid(lli->lli_jobid); | |
1171 | } else if (io->ci_type == CIT_SETATTR) { | |
1172 | if (!cl_io_is_trunc(io)) | |
1173 | io->ci_lockreq = CILR_MANDATORY; | |
1174 | } | |
1175 | ||
1176 | /* ignore layout change for generic CIT_MISC but not for glimpse. | |
1177 | * io context for glimpse must set ci_verify_layout to true, | |
1178 | * see cl_glimpse_size0() for details. */ | |
1179 | if (io->ci_type == CIT_MISC && !io->ci_verify_layout) | |
1180 | io->ci_ignore_layout = 1; | |
1181 | ||
1182 | /* Enqueue layout lock and get layout version. We need to do this | |
1183 | * even for operations requiring to open file, such as read and write, | |
1184 | * because it might not grant layout lock in IT_OPEN. */ | |
65fb55d1 | 1185 | if (result == 0 && !io->ci_ignore_layout) { |
d7e09d03 | 1186 | result = ll_layout_refresh(inode, &cio->cui_layout_gen); |
65fb55d1 NY |
1187 | if (result == -ENOENT) |
1188 | /* If the inode on MDS has been removed, but the objects | |
1189 | * on OSTs haven't been destroyed (async unlink), layout | |
d0a0acc3 | 1190 | * fetch will return -ENOENT, we'd ignore this error |
65fb55d1 NY |
1191 | * and continue with dirty flush. LU-3230. */ |
1192 | result = 0; | |
1193 | if (result < 0) | |
1194 | CERROR("%s: refresh file layout " DFID " error %d.\n", | |
1195 | ll_get_fsname(inode->i_sb, NULL, 0), | |
1196 | PFID(lu_object_fid(&obj->co_lu)), result); | |
1197 | } | |
d7e09d03 | 1198 | |
0a3bdb00 | 1199 | return result; |
d7e09d03 PT |
1200 | } |
1201 | ||
1202 | static struct vvp_io *cl2vvp_io(const struct lu_env *env, | |
1203 | const struct cl_io_slice *slice) | |
1204 | { | |
d0a0acc3 | 1205 | /* Calling just for assertion */ |
d7e09d03 PT |
1206 | cl2ccc_io(env, slice); |
1207 | return vvp_env_io(env); | |
1208 | } |