4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
27 * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #include <linux/kernel.h>
39 #include <linux/string.h>
40 #include <linux/stat.h>
41 #include <linux/errno.h>
42 #include <linux/unistd.h>
43 #include <asm/uaccess.h>
46 #include <linux/pagemap.h>
48 #define DEBUG_SUBSYSTEM S_LLITE
50 #include "../include/lustre_lite.h"
51 #include "llite_internal.h"
52 #include "../include/linux/lustre_compat25.h"
54 static const struct vm_operations_struct ll_file_vm_ops;
56 void policy_from_vma(ldlm_policy_data_t *policy,
57 struct vm_area_struct *vma, unsigned long addr,
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) |
66 struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
69 struct vm_area_struct *vma, *ret = NULL;
71 /* mmap_sem must have been held by caller. */
72 LASSERT(!down_write_trylock(&mm->mmap_sem));
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) {
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.
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.
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)
102 struct file *file = vma->vm_file;
103 struct inode *inode = file->f_dentry->d_inode;
105 struct cl_fault_io *fio;
110 if (ll_file_nolock(file))
111 return ERR_PTR(-EOPNOTSUPP);
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
120 env = cl_env_nested_get(nest);
122 return ERR_PTR(-EINVAL);
126 io = ccc_env_thread_io(env);
127 io->ci_obj = ll_i2info(inode)->lli_clob;
128 LASSERT(io->ci_obj != NULL);
130 fio = &io->u.ci_fault;
131 fio->ft_index = index;
132 fio->ft_executable = vma->vm_flags&VM_EXEC;
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.
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;
144 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
145 fio->ft_index, fio->ft_executable);
147 rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
149 struct ccc_io *cio = ccc_env_io(env);
150 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
152 LASSERT(cio->cui_cl.cis_io == io);
154 /* mmap lock must be MANDATORY it has to cache
156 io->ci_lockreq = CILR_MANDATORY;
161 cl_env_nested_put(nest, env);
168 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
169 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
175 struct cl_env_nest nest;
179 struct ll_inode_info *lli;
181 LASSERT(vmpage != NULL);
183 io = ll_fault_io_init(vma, &env, &nest, vmpage->index, NULL);
185 GOTO(out, result = PTR_ERR(io));
187 result = io->ci_result;
189 GOTO(out_io, result);
191 io->u.ci_fault.ft_mkwrite = 1;
192 io->u.ci_fault.ft_writable = 1;
194 vio = vvp_env_io(env);
195 vio->u.fault.ft_vma = vma;
196 vio->u.fault.ft_vmpage = vmpage;
198 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
200 /* we grab lli_trunc_sem to exclude truncate case.
201 * Otherwise, we could add dirty pages into osc cache
202 * while truncate is on-going. */
203 inode = ccc_object_inode(io->ci_obj);
204 lli = ll_i2info(inode);
205 down_read(&lli->lli_trunc_sem);
207 result = cl_io_loop(env, io);
209 up_read(&lli->lli_trunc_sem);
211 cfs_restore_sigs(set);
214 struct inode *inode = vma->vm_file->f_dentry->d_inode;
215 struct ll_inode_info *lli = ll_i2info(inode);
218 if (vmpage->mapping == NULL) {
221 /* page was truncated and lock was cancelled, return
222 * ENODATA so that VM_FAULT_NOPAGE will be returned
223 * to handle_mm_fault(). */
226 } else if (!PageDirty(vmpage)) {
227 /* race, the page has been cleaned by ptlrpcd after
228 * it was unlocked, it has to be added into dirty
229 * cache again otherwise this soon-to-dirty page won't
230 * consume any grants, even worse if this page is being
231 * transferred because it will break RPC checksum.
235 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
236 "been written out, retry.\n",
237 vmpage, vmpage->index);
244 spin_lock(&lli->lli_lock);
245 lli->lli_flags |= LLIF_DATA_MODIFIED;
246 spin_unlock(&lli->lli_lock);
252 cl_env_nested_put(&nest, env);
254 CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
255 LASSERT(ergo(result == 0, PageLocked(vmpage)));
262 static inline int to_fault_error(int result)
266 result = VM_FAULT_LOCKED;
269 result = VM_FAULT_NOPAGE;
272 result = VM_FAULT_OOM;
275 result = VM_FAULT_SIGBUS;
282 * Lustre implementation of a vm_operations_struct::fault() method, called by
283 * VM to server page fault (both in kernel and user space).
285 * \param vma - is virtual area struct related to page fault
286 * \param vmf - structure which describe type and address where hit fault
288 * \return allocated and filled _locked_ page for address
289 * \retval VM_FAULT_ERROR on general error
290 * \retval NOPAGE_OOM not have memory for allocate new page
292 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
296 struct vvp_io *vio = NULL;
298 unsigned long ra_flags;
299 struct cl_env_nest nest;
303 io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags);
305 return to_fault_error(PTR_ERR(io));
307 result = io->ci_result;
309 vio = vvp_env_io(env);
310 vio->u.fault.ft_vma = vma;
311 vio->u.fault.ft_vmpage = NULL;
312 vio->u.fault.fault.ft_vmf = vmf;
313 vio->u.fault.fault.ft_flags = 0;
314 vio->u.fault.fault.ft_flags_valid = 0;
316 result = cl_io_loop(env, io);
318 /* ft_flags are only valid if we reached
319 * the call to filemap_fault */
320 if (vio->u.fault.fault.ft_flags_valid)
321 fault_ret = vio->u.fault.fault.ft_flags;
323 vmpage = vio->u.fault.ft_vmpage;
324 if (result != 0 && vmpage != NULL) {
325 page_cache_release(vmpage);
330 cl_env_nested_put(&nest, env);
332 vma->vm_flags |= ra_flags;
333 if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
334 fault_ret |= to_fault_error(result);
336 CDEBUG(D_MMAP, "%s fault %d/%d\n",
337 current->comm, fault_ret, result);
341 static int ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
344 bool printed = false;
348 /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
349 * so that it can be killed by admin but not cause segfault by
351 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
354 result = ll_fault0(vma, vmf);
355 LASSERT(!(result & VM_FAULT_LOCKED));
357 struct page *vmpage = vmf->page;
359 /* check if this page has been truncated */
361 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
363 page_cache_release(vmpage);
366 if (!printed && ++count > 16) {
367 CWARN("the page is under heavy contention,"
368 "maybe your app(%s) needs revising :-)\n",
376 result = VM_FAULT_LOCKED;
378 cfs_restore_sigs(set);
382 static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
385 bool printed = false;
391 result = ll_page_mkwrite0(vma, vmf->page, &retry);
393 if (!printed && ++count > 16) {
394 CWARN("app(%s): the page %lu of file %lu is under heavy"
396 current->comm, vmf->pgoff,
397 vma->vm_file->f_dentry->d_inode->i_ino);
404 LASSERT(PageLocked(vmf->page));
405 result = VM_FAULT_LOCKED;
409 result = VM_FAULT_NOPAGE;
412 result = VM_FAULT_OOM;
415 result = VM_FAULT_RETRY;
418 result = VM_FAULT_SIGBUS;
426 * To avoid cancel the locks covering mmapped region for lock cache pressure,
427 * we track the mapped vma count in ccc_object::cob_mmap_cnt.
429 static void ll_vm_open(struct vm_area_struct * vma)
431 struct inode *inode = vma->vm_file->f_dentry->d_inode;
432 struct ccc_object *vob = cl_inode2ccc(inode);
434 LASSERT(vma->vm_file);
435 LASSERT(atomic_read(&vob->cob_mmap_cnt) >= 0);
436 atomic_inc(&vob->cob_mmap_cnt);
440 * Dual to ll_vm_open().
442 static void ll_vm_close(struct vm_area_struct *vma)
444 struct inode *inode = vma->vm_file->f_dentry->d_inode;
445 struct ccc_object *vob = cl_inode2ccc(inode);
447 LASSERT(vma->vm_file);
448 atomic_dec(&vob->cob_mmap_cnt);
449 LASSERT(atomic_read(&vob->cob_mmap_cnt) >= 0);
452 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
453 * nopage's reference passing to the pte */
454 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
458 LASSERTF(last > first, "last %llu first %llu\n", last, first);
459 if (mapping_mapped(mapping)) {
461 unmap_mapping_range(mapping, first + PAGE_CACHE_SIZE - 1,
462 last - first + 1, 0);
468 static const struct vm_operations_struct ll_file_vm_ops = {
470 .page_mkwrite = ll_page_mkwrite,
472 .close = ll_vm_close,
475 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
477 struct inode *inode = file->f_dentry->d_inode;
480 if (ll_file_nolock(file))
483 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
484 rc = generic_file_mmap(file, vma);
486 vma->vm_ops = &ll_file_vm_ops;
487 vma->vm_ops->open(vma);
488 /* update the inode's size and mtime */
489 rc = ll_glimpse_size(inode);