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 *
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
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>
43 #include <linux/uaccess.h>
44
45 #include <linux/fs.h>
46 #include <linux/pagemap.h>
47
48 #define DEBUG_SUBSYSTEM S_LLITE
49
50 #include "../include/lustre_lite.h"
51 #include "llite_internal.h"
52 #include "../include/linux/lustre_compat25.h"
53
54 static const struct vm_operations_struct ll_file_vm_ops;
55
policy_from_vma(ldlm_policy_data_t * policy,struct vm_area_struct * vma,unsigned long addr,size_t count)56 void policy_from_vma(ldlm_policy_data_t *policy,
57 struct vm_area_struct *vma, unsigned long addr,
58 size_t count)
59 {
60 policy->l_extent.start = ((addr - vma->vm_start) & CFS_PAGE_MASK) +
61 (vma->vm_pgoff << PAGE_CACHE_SHIFT);
62 policy->l_extent.end = (policy->l_extent.start + count - 1) |
63 ~CFS_PAGE_MASK;
64 }
65
our_vma(struct mm_struct * mm,unsigned long addr,size_t count)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;
70
71 /* mmap_sem must have been held by caller. */
72 LASSERT(!down_write_trylock(&mm->mmap_sem));
73
74 for (vma = find_vma(mm, addr);
75 vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
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 }
82 return ret;
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 */
97 static struct cl_io *
ll_fault_io_init(struct vm_area_struct * vma,struct lu_env ** env_ret,struct cl_env_nest * nest,pgoff_t index,unsigned long * ra_flags)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)
101 {
102 struct file *file = vma->vm_file;
103 struct inode *inode = file_inode(file);
104 struct cl_io *io;
105 struct cl_fault_io *fio;
106 struct lu_env *env;
107 int rc;
108
109 *env_ret = NULL;
110 if (ll_file_nolock(file))
111 return ERR_PTR(-EOPNOTSUPP);
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))
122 return ERR_PTR(-EINVAL);
123
124 *env_ret = env;
125
126 io = ccc_env_thread_io(env);
127 io->ci_obj = ll_i2info(inode)->lli_clob;
128 LASSERT(io->ci_obj != NULL);
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 */
139 if (ra_flags != NULL)
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
147 rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
148 if (rc == 0) {
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
154 /* mmap lock must be MANDATORY it has to cache
155 * pages. */
156 io->ci_lockreq = CILR_MANDATORY;
157 cio->cui_fd = fd;
158 } else {
159 LASSERT(rc < 0);
160 cl_io_fini(env, io);
161 cl_env_nested_put(nest, env);
162 io = ERR_PTR(rc);
163 }
164
165 return io;
166 }
167
168 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
ll_page_mkwrite0(struct vm_area_struct * vma,struct page * vmpage,bool * retry)169 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
170 bool *retry)
171 {
172 struct lu_env *env;
173 struct cl_io *io;
174 struct vvp_io *vio;
175 struct cl_env_nest nest;
176 int result;
177 sigset_t set;
178 struct inode *inode;
179 struct ll_inode_info *lli;
180
181 LASSERT(vmpage != NULL);
182
183 io = ll_fault_io_init(vma, &env, &nest, vmpage->index, NULL);
184 if (IS_ERR(io)) {
185 result = PTR_ERR(io);
186 goto out;
187 }
188
189 result = io->ci_result;
190 if (result < 0)
191 goto out_io;
192
193 io->u.ci_fault.ft_mkwrite = 1;
194 io->u.ci_fault.ft_writable = 1;
195
196 vio = vvp_env_io(env);
197 vio->u.fault.ft_vma = vma;
198 vio->u.fault.ft_vmpage = vmpage;
199
200 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
201
202 /* we grab lli_trunc_sem to exclude truncate case.
203 * Otherwise, we could add dirty pages into osc cache
204 * while truncate is on-going. */
205 inode = ccc_object_inode(io->ci_obj);
206 lli = ll_i2info(inode);
207 down_read(&lli->lli_trunc_sem);
208
209 result = cl_io_loop(env, io);
210
211 up_read(&lli->lli_trunc_sem);
212
213 cfs_restore_sigs(set);
214
215 if (result == 0) {
216 struct inode *inode = file_inode(vma->vm_file);
217 struct ll_inode_info *lli = ll_i2info(inode);
218
219 lock_page(vmpage);
220 if (vmpage->mapping == NULL) {
221 unlock_page(vmpage);
222
223 /* page was truncated and lock was cancelled, return
224 * ENODATA so that VM_FAULT_NOPAGE will be returned
225 * to handle_mm_fault(). */
226 if (result == 0)
227 result = -ENODATA;
228 } else if (!PageDirty(vmpage)) {
229 /* race, the page has been cleaned by ptlrpcd after
230 * it was unlocked, it has to be added into dirty
231 * cache again otherwise this soon-to-dirty page won't
232 * consume any grants, even worse if this page is being
233 * transferred because it will break RPC checksum.
234 */
235 unlock_page(vmpage);
236
237 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has been written out, retry.\n",
238 vmpage, vmpage->index);
239
240 *retry = true;
241 result = -EAGAIN;
242 }
243
244 if (result == 0) {
245 spin_lock(&lli->lli_lock);
246 lli->lli_flags |= LLIF_DATA_MODIFIED;
247 spin_unlock(&lli->lli_lock);
248 }
249 }
250
251 out_io:
252 cl_io_fini(env, io);
253 cl_env_nested_put(&nest, env);
254 out:
255 CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
256 LASSERT(ergo(result == 0, PageLocked(vmpage)));
257
258 return result;
259 }
260
to_fault_error(int result)261 static inline int to_fault_error(int result)
262 {
263 switch (result) {
264 case 0:
265 result = VM_FAULT_LOCKED;
266 break;
267 case -EFAULT:
268 result = VM_FAULT_NOPAGE;
269 break;
270 case -ENOMEM:
271 result = VM_FAULT_OOM;
272 break;
273 default:
274 result = VM_FAULT_SIGBUS;
275 break;
276 }
277 return result;
278 }
279
280 /**
281 * Lustre implementation of a vm_operations_struct::fault() method, called by
282 * VM to server page fault (both in kernel and user space).
283 *
284 * \param vma - is virtual area struct related to page fault
285 * \param vmf - structure which describe type and address where hit fault
286 *
287 * \return allocated and filled _locked_ page for address
288 * \retval VM_FAULT_ERROR on general error
289 * \retval NOPAGE_OOM not have memory for allocate new page
290 */
ll_fault0(struct vm_area_struct * vma,struct vm_fault * vmf)291 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
292 {
293 struct lu_env *env;
294 struct cl_io *io;
295 struct vvp_io *vio = NULL;
296 struct page *vmpage;
297 unsigned long ra_flags;
298 struct cl_env_nest nest;
299 int result;
300 int fault_ret = 0;
301
302 io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags);
303 if (IS_ERR(io))
304 return to_fault_error(PTR_ERR(io));
305
306 result = io->ci_result;
307 if (result == 0) {
308 vio = vvp_env_io(env);
309 vio->u.fault.ft_vma = vma;
310 vio->u.fault.ft_vmpage = NULL;
311 vio->u.fault.fault.ft_vmf = vmf;
312 vio->u.fault.fault.ft_flags = 0;
313 vio->u.fault.fault.ft_flags_valid = false;
314
315 result = cl_io_loop(env, io);
316
317 /* ft_flags are only valid if we reached
318 * the call to filemap_fault */
319 if (vio->u.fault.fault.ft_flags_valid)
320 fault_ret = vio->u.fault.fault.ft_flags;
321
322 vmpage = vio->u.fault.ft_vmpage;
323 if (result != 0 && vmpage != NULL) {
324 page_cache_release(vmpage);
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);
337 return fault_ret;
338 }
339
ll_fault(struct vm_area_struct * vma,struct vm_fault * vmf)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
347 /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
348 * so that it can be killed by admin but not cause segfault by
349 * other signals. */
350 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
351
352 restart:
353 result = ll_fault0(vma, vmf);
354 LASSERT(!(result & VM_FAULT_LOCKED));
355 if (result == 0) {
356 struct page *vmpage = vmf->page;
357
358 /* check if this page has been truncated */
359 lock_page(vmpage);
360 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
361 unlock_page(vmpage);
362 page_cache_release(vmpage);
363 vmf->page = NULL;
364
365 if (!printed && ++count > 16) {
366 CWARN("the page is under heavy contention, maybe your app(%s) needs revising :-)\n",
367 current->comm);
368 printed = true;
369 }
370
371 goto restart;
372 }
373
374 result = VM_FAULT_LOCKED;
375 }
376 cfs_restore_sigs(set);
377 return result;
378 }
379
ll_page_mkwrite(struct vm_area_struct * vma,struct vm_fault * vmf)380 static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
381 {
382 int count = 0;
383 bool printed = false;
384 bool retry;
385 int result;
386
387 do {
388 retry = false;
389 result = ll_page_mkwrite0(vma, vmf->page, &retry);
390
391 if (!printed && ++count > 16) {
392 CWARN("app(%s): the page %lu of file %lu is under heavy contention.\n",
393 current->comm, vmf->pgoff,
394 file_inode(vma->vm_file)->i_ino);
395 printed = true;
396 }
397 } while (retry);
398
399 switch (result) {
400 case 0:
401 LASSERT(PageLocked(vmf->page));
402 result = VM_FAULT_LOCKED;
403 break;
404 case -ENODATA:
405 case -EAGAIN:
406 case -EFAULT:
407 result = VM_FAULT_NOPAGE;
408 break;
409 case -ENOMEM:
410 result = VM_FAULT_OOM;
411 break;
412 default:
413 result = VM_FAULT_SIGBUS;
414 break;
415 }
416
417 return result;
418 }
419
420 /**
421 * To avoid cancel the locks covering mmapped region for lock cache pressure,
422 * we track the mapped vma count in ccc_object::cob_mmap_cnt.
423 */
ll_vm_open(struct vm_area_struct * vma)424 static void ll_vm_open(struct vm_area_struct *vma)
425 {
426 struct inode *inode = file_inode(vma->vm_file);
427 struct ccc_object *vob = cl_inode2ccc(inode);
428
429 LASSERT(vma->vm_file);
430 LASSERT(atomic_read(&vob->cob_mmap_cnt) >= 0);
431 atomic_inc(&vob->cob_mmap_cnt);
432 }
433
434 /**
435 * Dual to ll_vm_open().
436 */
ll_vm_close(struct vm_area_struct * vma)437 static void ll_vm_close(struct vm_area_struct *vma)
438 {
439 struct inode *inode = file_inode(vma->vm_file);
440 struct ccc_object *vob = cl_inode2ccc(inode);
441
442 LASSERT(vma->vm_file);
443 atomic_dec(&vob->cob_mmap_cnt);
444 LASSERT(atomic_read(&vob->cob_mmap_cnt) >= 0);
445 }
446
447 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
448 * nopage's reference passing to the pte */
ll_teardown_mmaps(struct address_space * mapping,__u64 first,__u64 last)449 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
450 {
451 int rc = -ENOENT;
452
453 LASSERTF(last > first, "last %llu first %llu\n", last, first);
454 if (mapping_mapped(mapping)) {
455 rc = 0;
456 unmap_mapping_range(mapping, first + PAGE_CACHE_SIZE - 1,
457 last - first + 1, 0);
458 }
459
460 return rc;
461 }
462
463 static const struct vm_operations_struct ll_file_vm_ops = {
464 .fault = ll_fault,
465 .page_mkwrite = ll_page_mkwrite,
466 .open = ll_vm_open,
467 .close = ll_vm_close,
468 };
469
ll_file_mmap(struct file * file,struct vm_area_struct * vma)470 int ll_file_mmap(struct file *file, struct vm_area_struct *vma)
471 {
472 struct inode *inode = file_inode(file);
473 int rc;
474
475 if (ll_file_nolock(file))
476 return -EOPNOTSUPP;
477
478 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
479 rc = generic_file_mmap(file, vma);
480 if (rc == 0) {
481 vma->vm_ops = &ll_file_vm_ops;
482 vma->vm_ops->open(vma);
483 /* update the inode's size and mtime */
484 rc = ll_glimpse_size(inode);
485 }
486
487 return rc;
488 }
489