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) 2004, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
1dc563a6 | 30 | * Copyright (c) 2011, 2015, Intel Corporation. |
d7e09d03 PT |
31 | */ |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | */ | |
36 | ||
37 | #include <linux/kernel.h> | |
38 | #include <linux/mm.h> | |
39 | #include <linux/string.h> | |
40 | #include <linux/stat.h> | |
41 | #include <linux/errno.h> | |
42 | #include <linux/unistd.h> | |
e8fd99fd | 43 | #include <linux/uaccess.h> |
d7e09d03 PT |
44 | |
45 | #include <linux/fs.h> | |
d7e09d03 PT |
46 | #include <linux/pagemap.h> |
47 | ||
48 | #define DEBUG_SUBSYSTEM S_LLITE | |
49 | ||
67a235f5 | 50 | #include "../include/lustre_lite.h" |
d7e09d03 | 51 | #include "llite_internal.h" |
67a235f5 | 52 | #include "../include/linux/lustre_compat25.h" |
d7e09d03 | 53 | |
2d95f10e | 54 | static const struct vm_operations_struct ll_file_vm_ops; |
d7e09d03 PT |
55 | |
56 | void policy_from_vma(ldlm_policy_data_t *policy, | |
e15ba45d OD |
57 | struct vm_area_struct *vma, unsigned long addr, |
58 | size_t count) | |
d7e09d03 PT |
59 | { |
60 | policy->l_extent.start = ((addr - vma->vm_start) & CFS_PAGE_MASK) + | |
09cbfeaf | 61 | (vma->vm_pgoff << PAGE_SHIFT); |
d7e09d03 PT |
62 | policy->l_extent.end = (policy->l_extent.start + count - 1) | |
63 | ~CFS_PAGE_MASK; | |
64 | } | |
65 | ||
66 | struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr, | |
67 | size_t count) | |
68 | { | |
69 | struct vm_area_struct *vma, *ret = NULL; | |
d7e09d03 PT |
70 | |
71 | /* mmap_sem must have been held by caller. */ | |
72 | LASSERT(!down_write_trylock(&mm->mmap_sem)); | |
73 | ||
a58a38ac | 74 | for (vma = find_vma(mm, addr); |
6e16818b | 75 | vma && vma->vm_start < (addr + count); vma = vma->vm_next) { |
d7e09d03 PT |
76 | if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops && |
77 | vma->vm_flags & VM_SHARED) { | |
78 | ret = vma; | |
79 | break; | |
80 | } | |
81 | } | |
0a3bdb00 | 82 | return ret; |
d7e09d03 PT |
83 | } |
84 | ||
85 | /** | |
86 | * API independent part for page fault initialization. | |
87 | * \param vma - virtual memory area addressed to page fault | |
88 | * \param env - corespondent lu_env to processing | |
89 | * \param nest - nested level | |
90 | * \param index - page index corespondent to fault. | |
91 | * \parm ra_flags - vma readahead flags. | |
92 | * | |
93 | * \return allocated and initialized env for fault operation. | |
94 | * \retval EINVAL if env can't allocated | |
95 | * \return other error codes from cl_io_init. | |
96 | */ | |
2d95f10e JH |
97 | static struct cl_io * |
98 | ll_fault_io_init(struct vm_area_struct *vma, struct lu_env **env_ret, | |
99 | struct cl_env_nest *nest, pgoff_t index, | |
100 | unsigned long *ra_flags) | |
d7e09d03 | 101 | { |
8a48df70 | 102 | struct file *file = vma->vm_file; |
2a8a3597 | 103 | struct inode *inode = file_inode(file); |
8a48df70 JH |
104 | struct cl_io *io; |
105 | struct cl_fault_io *fio; | |
106 | struct lu_env *env; | |
107 | int rc; | |
d7e09d03 PT |
108 | |
109 | *env_ret = NULL; | |
110 | if (ll_file_nolock(file)) | |
0a3bdb00 | 111 | return ERR_PTR(-EOPNOTSUPP); |
d7e09d03 PT |
112 | |
113 | /* | |
114 | * page fault can be called when lustre IO is | |
115 | * already active for the current thread, e.g., when doing read/write | |
116 | * against user level buffer mapped from Lustre buffer. To avoid | |
117 | * stomping on existing context, optionally force an allocation of a new | |
118 | * one. | |
119 | */ | |
120 | env = cl_env_nested_get(nest); | |
121 | if (IS_ERR(env)) | |
defa220f | 122 | return ERR_PTR(-EINVAL); |
d7e09d03 PT |
123 | |
124 | *env_ret = env; | |
125 | ||
126 | io = ccc_env_thread_io(env); | |
127 | io->ci_obj = ll_i2info(inode)->lli_clob; | |
6e16818b | 128 | LASSERT(io->ci_obj); |
d7e09d03 PT |
129 | |
130 | fio = &io->u.ci_fault; | |
131 | fio->ft_index = index; | |
132 | fio->ft_executable = vma->vm_flags&VM_EXEC; | |
133 | ||
134 | /* | |
135 | * disable VM_SEQ_READ and use VM_RAND_READ to make sure that | |
136 | * the kernel will not read other pages not covered by ldlm in | |
137 | * filemap_nopage. we do our readahead in ll_readpage. | |
138 | */ | |
6e16818b | 139 | if (ra_flags) |
d7e09d03 PT |
140 | *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ); |
141 | vma->vm_flags &= ~VM_SEQ_READ; | |
142 | vma->vm_flags |= VM_RAND_READ; | |
143 | ||
144 | CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags, | |
145 | fio->ft_index, fio->ft_executable); | |
146 | ||
8a48df70 JH |
147 | rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj); |
148 | if (rc == 0) { | |
d7e09d03 PT |
149 | struct ccc_io *cio = ccc_env_io(env); |
150 | struct ll_file_data *fd = LUSTRE_FPRIVATE(file); | |
151 | ||
152 | LASSERT(cio->cui_cl.cis_io == io); | |
153 | ||
c0894c6c | 154 | /* mmap lock must be MANDATORY it has to cache pages. */ |
d7e09d03 | 155 | io->ci_lockreq = CILR_MANDATORY; |
8a48df70 JH |
156 | cio->cui_fd = fd; |
157 | } else { | |
158 | LASSERT(rc < 0); | |
159 | cl_io_fini(env, io); | |
160 | cl_env_nested_put(nest, env); | |
161 | io = ERR_PTR(rc); | |
d7e09d03 PT |
162 | } |
163 | ||
164 | return io; | |
165 | } | |
166 | ||
167 | /* Sharing code of page_mkwrite method for rhel5 and rhel6 */ | |
168 | static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage, | |
169 | bool *retry) | |
170 | { | |
171 | struct lu_env *env; | |
172 | struct cl_io *io; | |
173 | struct vvp_io *vio; | |
174 | struct cl_env_nest nest; | |
175 | int result; | |
176 | sigset_t set; | |
177 | struct inode *inode; | |
178 | struct ll_inode_info *lli; | |
d7e09d03 | 179 | |
d7e09d03 | 180 | io = ll_fault_io_init(vma, &env, &nest, vmpage->index, NULL); |
34e1f2bb JL |
181 | if (IS_ERR(io)) { |
182 | result = PTR_ERR(io); | |
183 | goto out; | |
184 | } | |
d7e09d03 PT |
185 | |
186 | result = io->ci_result; | |
187 | if (result < 0) | |
34e1f2bb | 188 | goto out_io; |
d7e09d03 PT |
189 | |
190 | io->u.ci_fault.ft_mkwrite = 1; | |
191 | io->u.ci_fault.ft_writable = 1; | |
192 | ||
193 | vio = vvp_env_io(env); | |
194 | vio->u.fault.ft_vma = vma; | |
195 | vio->u.fault.ft_vmpage = vmpage; | |
196 | ||
197 | set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM)); | |
198 | ||
199 | /* we grab lli_trunc_sem to exclude truncate case. | |
200 | * Otherwise, we could add dirty pages into osc cache | |
c0894c6c OD |
201 | * while truncate is on-going. |
202 | */ | |
d7e09d03 PT |
203 | inode = ccc_object_inode(io->ci_obj); |
204 | lli = ll_i2info(inode); | |
205 | down_read(&lli->lli_trunc_sem); | |
206 | ||
207 | result = cl_io_loop(env, io); | |
208 | ||
209 | up_read(&lli->lli_trunc_sem); | |
210 | ||
211 | cfs_restore_sigs(set); | |
212 | ||
213 | if (result == 0) { | |
2a8a3597 | 214 | struct inode *inode = file_inode(vma->vm_file); |
d7e09d03 PT |
215 | struct ll_inode_info *lli = ll_i2info(inode); |
216 | ||
217 | lock_page(vmpage); | |
6e16818b | 218 | if (!vmpage->mapping) { |
d7e09d03 PT |
219 | unlock_page(vmpage); |
220 | ||
221 | /* page was truncated and lock was cancelled, return | |
222 | * ENODATA so that VM_FAULT_NOPAGE will be returned | |
c0894c6c OD |
223 | * to handle_mm_fault(). |
224 | */ | |
d7e09d03 PT |
225 | if (result == 0) |
226 | result = -ENODATA; | |
227 | } else if (!PageDirty(vmpage)) { | |
228 | /* race, the page has been cleaned by ptlrpcd after | |
229 | * it was unlocked, it has to be added into dirty | |
230 | * cache again otherwise this soon-to-dirty page won't | |
231 | * consume any grants, even worse if this page is being | |
232 | * transferred because it will break RPC checksum. | |
233 | */ | |
234 | unlock_page(vmpage); | |
235 | ||
2d00bd17 | 236 | CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has been written out, retry.\n", |
d7e09d03 PT |
237 | vmpage, vmpage->index); |
238 | ||
239 | *retry = true; | |
240 | result = -EAGAIN; | |
241 | } | |
242 | ||
243 | if (result == 0) { | |
244 | spin_lock(&lli->lli_lock); | |
245 | lli->lli_flags |= LLIF_DATA_MODIFIED; | |
246 | spin_unlock(&lli->lli_lock); | |
247 | } | |
248 | } | |
d7e09d03 | 249 | |
8a48df70 | 250 | out_io: |
d7e09d03 PT |
251 | cl_io_fini(env, io); |
252 | cl_env_nested_put(&nest, env); | |
8a48df70 | 253 | out: |
d7e09d03 | 254 | CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result); |
d7e09d03 | 255 | LASSERT(ergo(result == 0, PageLocked(vmpage))); |
8a48df70 JH |
256 | |
257 | return result; | |
d7e09d03 PT |
258 | } |
259 | ||
d7e09d03 PT |
260 | static inline int to_fault_error(int result) |
261 | { | |
a58a38ac | 262 | switch (result) { |
d7e09d03 PT |
263 | case 0: |
264 | result = VM_FAULT_LOCKED; | |
265 | break; | |
266 | case -EFAULT: | |
267 | result = VM_FAULT_NOPAGE; | |
268 | break; | |
269 | case -ENOMEM: | |
270 | result = VM_FAULT_OOM; | |
271 | break; | |
272 | default: | |
273 | result = VM_FAULT_SIGBUS; | |
274 | break; | |
275 | } | |
276 | return result; | |
277 | } | |
278 | ||
279 | /** | |
280 | * Lustre implementation of a vm_operations_struct::fault() method, called by | |
281 | * VM to server page fault (both in kernel and user space). | |
282 | * | |
d0a0acc3 | 283 | * \param vma - is virtual area struct related to page fault |
d7e09d03 PT |
284 | * \param vmf - structure which describe type and address where hit fault |
285 | * | |
286 | * \return allocated and filled _locked_ page for address | |
287 | * \retval VM_FAULT_ERROR on general error | |
288 | * \retval NOPAGE_OOM not have memory for allocate new page | |
289 | */ | |
290 | static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf) | |
291 | { | |
292 | struct lu_env *env; | |
293 | struct cl_io *io; | |
294 | struct vvp_io *vio = NULL; | |
295 | struct page *vmpage; | |
296 | unsigned long ra_flags; | |
297 | struct cl_env_nest nest; | |
298 | int result; | |
299 | int fault_ret = 0; | |
d7e09d03 PT |
300 | |
301 | io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags); | |
302 | if (IS_ERR(io)) | |
0a3bdb00 | 303 | return to_fault_error(PTR_ERR(io)); |
d7e09d03 PT |
304 | |
305 | result = io->ci_result; | |
306 | if (result == 0) { | |
307 | vio = vvp_env_io(env); | |
308 | vio->u.fault.ft_vma = vma; | |
309 | vio->u.fault.ft_vmpage = NULL; | |
310 | vio->u.fault.fault.ft_vmf = vmf; | |
6aa51072 | 311 | vio->u.fault.fault.ft_flags = 0; |
19cd2297 | 312 | vio->u.fault.fault.ft_flags_valid = false; |
d7e09d03 PT |
313 | |
314 | result = cl_io_loop(env, io); | |
315 | ||
6aa51072 | 316 | /* ft_flags are only valid if we reached |
c0894c6c OD |
317 | * the call to filemap_fault |
318 | */ | |
6aa51072 PC |
319 | if (vio->u.fault.fault.ft_flags_valid) |
320 | fault_ret = vio->u.fault.fault.ft_flags; | |
321 | ||
d7e09d03 | 322 | vmpage = vio->u.fault.ft_vmpage; |
6e16818b | 323 | if (result != 0 && vmpage) { |
09cbfeaf | 324 | put_page(vmpage); |
d7e09d03 PT |
325 | vmf->page = NULL; |
326 | } | |
327 | } | |
328 | cl_io_fini(env, io); | |
329 | cl_env_nested_put(&nest, env); | |
330 | ||
331 | vma->vm_flags |= ra_flags; | |
332 | if (result != 0 && !(fault_ret & VM_FAULT_RETRY)) | |
333 | fault_ret |= to_fault_error(result); | |
334 | ||
335 | CDEBUG(D_MMAP, "%s fault %d/%d\n", | |
336 | current->comm, fault_ret, result); | |
0a3bdb00 | 337 | return fault_ret; |
d7e09d03 PT |
338 | } |
339 | ||
340 | static int ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf) | |
341 | { | |
342 | int count = 0; | |
343 | bool printed = false; | |
344 | int result; | |
345 | sigset_t set; | |
346 | ||
c0894c6c | 347 | /* Only SIGKILL and SIGTERM are allowed for fault/nopage/mkwrite |
d7e09d03 | 348 | * so that it can be killed by admin but not cause segfault by |
c0894c6c OD |
349 | * other signals. |
350 | */ | |
d7e09d03 PT |
351 | set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM)); |
352 | ||
353 | restart: | |
354 | result = ll_fault0(vma, vmf); | |
355 | LASSERT(!(result & VM_FAULT_LOCKED)); | |
356 | if (result == 0) { | |
357 | struct page *vmpage = vmf->page; | |
358 | ||
359 | /* check if this page has been truncated */ | |
360 | lock_page(vmpage); | |
6e16818b | 361 | if (unlikely(!vmpage->mapping)) { /* unlucky */ |
d7e09d03 | 362 | unlock_page(vmpage); |
09cbfeaf | 363 | put_page(vmpage); |
d7e09d03 PT |
364 | vmf->page = NULL; |
365 | ||
366 | if (!printed && ++count > 16) { | |
2d00bd17 | 367 | CWARN("the page is under heavy contention, maybe your app(%s) needs revising :-)\n", |
d7e09d03 PT |
368 | current->comm); |
369 | printed = true; | |
370 | } | |
371 | ||
372 | goto restart; | |
373 | } | |
374 | ||
34d1f637 | 375 | result = VM_FAULT_LOCKED; |
d7e09d03 PT |
376 | } |
377 | cfs_restore_sigs(set); | |
378 | return result; | |
379 | } | |
380 | ||
381 | static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf) | |
382 | { | |
383 | int count = 0; | |
384 | bool printed = false; | |
385 | bool retry; | |
386 | int result; | |
387 | ||
388 | do { | |
389 | retry = false; | |
390 | result = ll_page_mkwrite0(vma, vmf->page, &retry); | |
391 | ||
392 | if (!printed && ++count > 16) { | |
2d00bd17 | 393 | CWARN("app(%s): the page %lu of file %lu is under heavy contention.\n", |
d7e09d03 | 394 | current->comm, vmf->pgoff, |
2a8a3597 | 395 | file_inode(vma->vm_file)->i_ino); |
d7e09d03 PT |
396 | printed = true; |
397 | } | |
398 | } while (retry); | |
399 | ||
a58a38ac | 400 | switch (result) { |
d7e09d03 PT |
401 | case 0: |
402 | LASSERT(PageLocked(vmf->page)); | |
403 | result = VM_FAULT_LOCKED; | |
404 | break; | |
405 | case -ENODATA: | |
406 | case -EFAULT: | |
407 | result = VM_FAULT_NOPAGE; | |
408 | break; | |
409 | case -ENOMEM: | |
410 | result = VM_FAULT_OOM; | |
411 | break; | |
412 | case -EAGAIN: | |
413 | result = VM_FAULT_RETRY; | |
414 | break; | |
415 | default: | |
416 | result = VM_FAULT_SIGBUS; | |
417 | break; | |
418 | } | |
419 | ||
420 | return result; | |
421 | } | |
422 | ||
423 | /** | |
424 | * To avoid cancel the locks covering mmapped region for lock cache pressure, | |
425 | * we track the mapped vma count in ccc_object::cob_mmap_cnt. | |
426 | */ | |
aff9d8e8 | 427 | static void ll_vm_open(struct vm_area_struct *vma) |
d7e09d03 | 428 | { |
2a8a3597 | 429 | struct inode *inode = file_inode(vma->vm_file); |
d7e09d03 PT |
430 | struct ccc_object *vob = cl_inode2ccc(inode); |
431 | ||
d7e09d03 PT |
432 | LASSERT(vma->vm_file); |
433 | LASSERT(atomic_read(&vob->cob_mmap_cnt) >= 0); | |
434 | atomic_inc(&vob->cob_mmap_cnt); | |
d7e09d03 PT |
435 | } |
436 | ||
437 | /** | |
438 | * Dual to ll_vm_open(). | |
439 | */ | |
440 | static void ll_vm_close(struct vm_area_struct *vma) | |
441 | { | |
2a8a3597 | 442 | struct inode *inode = file_inode(vma->vm_file); |
d7e09d03 PT |
443 | struct ccc_object *vob = cl_inode2ccc(inode); |
444 | ||
d7e09d03 PT |
445 | LASSERT(vma->vm_file); |
446 | atomic_dec(&vob->cob_mmap_cnt); | |
447 | LASSERT(atomic_read(&vob->cob_mmap_cnt) >= 0); | |
d7e09d03 PT |
448 | } |
449 | ||
d7e09d03 | 450 | /* XXX put nice comment here. talk about __free_pte -> dirty pages and |
c0894c6c OD |
451 | * nopage's reference passing to the pte |
452 | */ | |
d7e09d03 PT |
453 | int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last) |
454 | { | |
455 | int rc = -ENOENT; | |
d7e09d03 | 456 | |
b0f5aad5 | 457 | LASSERTF(last > first, "last %llu first %llu\n", last, first); |
d7e09d03 PT |
458 | if (mapping_mapped(mapping)) { |
459 | rc = 0; | |
09cbfeaf | 460 | unmap_mapping_range(mapping, first + PAGE_SIZE - 1, |
d7e09d03 PT |
461 | last - first + 1, 0); |
462 | } | |
463 | ||
0a3bdb00 | 464 | return rc; |
d7e09d03 PT |
465 | } |
466 | ||
2d95f10e | 467 | static const struct vm_operations_struct ll_file_vm_ops = { |
d7e09d03 PT |
468 | .fault = ll_fault, |
469 | .page_mkwrite = ll_page_mkwrite, | |
470 | .open = ll_vm_open, | |
471 | .close = ll_vm_close, | |
472 | }; | |
473 | ||
aff9d8e8 | 474 | int ll_file_mmap(struct file *file, struct vm_area_struct *vma) |
d7e09d03 | 475 | { |
2a8a3597 | 476 | struct inode *inode = file_inode(file); |
d7e09d03 | 477 | int rc; |
d7e09d03 PT |
478 | |
479 | if (ll_file_nolock(file)) | |
0a3bdb00 | 480 | return -EOPNOTSUPP; |
d7e09d03 PT |
481 | |
482 | ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1); | |
483 | rc = generic_file_mmap(file, vma); | |
484 | if (rc == 0) { | |
485 | vma->vm_ops = &ll_file_vm_ops; | |
486 | vma->vm_ops->open(vma); | |
487 | /* update the inode's size and mtime */ | |
488 | rc = ll_glimpse_size(inode); | |
489 | } | |
490 | ||
0a3bdb00 | 491 | return rc; |
d7e09d03 | 492 | } |