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, 2015, 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 <linux/version.h>
44 #include <asm/uaccess.h>
47 #include <linux/stat.h>
48 #include <asm/uaccess.h>
50 #include <linux/pagemap.h>
52 #define DEBUG_SUBSYSTEM S_LLITE
54 #include "llite_internal.h"
55 #include <lustre_compat.h>
57 static const struct vm_operations_struct ll_file_vm_ops;
59 void policy_from_vma(union ldlm_policy_data *policy, struct vm_area_struct *vma,
60 unsigned long addr, size_t count)
62 policy->l_extent.start = ((addr - vma->vm_start) & PAGE_MASK) +
63 (vma->vm_pgoff << PAGE_CACHE_SHIFT);
64 policy->l_extent.end = (policy->l_extent.start + count - 1) |
68 struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
71 struct vm_area_struct *vma, *ret = NULL;
74 /* mmap_sem must have been held by caller. */
75 LASSERT(!down_write_trylock(&mm->mmap_sem));
77 for(vma = find_vma(mm, addr);
78 vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
79 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
80 vma->vm_flags & VM_SHARED) {
89 * API independent part for page fault initialization.
90 * \param env - corespondent lu_env to processing
91 * \param vma - virtual memory area addressed to page fault
92 * \param index - page index corespondent to fault.
93 * \parm ra_flags - vma readahead flags.
95 * \return error codes from cl_io_init.
98 ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma,
99 pgoff_t index, unsigned long *ra_flags)
101 struct file *file = vma->vm_file;
102 struct inode *inode = file->f_path.dentry->d_inode;
104 struct cl_fault_io *fio;
108 if (ll_file_nolock(file))
109 RETURN(ERR_PTR(-EOPNOTSUPP));
112 io = vvp_env_thread_io(env);
113 io->ci_obj = ll_i2info(inode)->lli_clob;
114 LASSERT(io->ci_obj != NULL);
116 fio = &io->u.ci_fault;
117 fio->ft_index = index;
118 fio->ft_executable = vma->vm_flags&VM_EXEC;
121 * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
122 * the kernel will not read other pages not covered by ldlm in
123 * filemap_nopage. we do our readahead in ll_readpage.
125 if (ra_flags != NULL)
126 *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
127 vma->vm_flags &= ~VM_SEQ_READ;
128 vma->vm_flags |= VM_RAND_READ;
130 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
131 fio->ft_index, fio->ft_executable);
133 rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
135 struct vvp_io *vio = vvp_env_io(env);
136 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
138 LASSERT(vio->vui_cl.cis_io == io);
140 /* mmap lock must be MANDATORY it has to cache
142 io->ci_lockreq = CILR_MANDATORY;
147 if (io->ci_need_restart)
156 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
157 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
167 struct ll_inode_info *lli;
170 LASSERT(vmpage != NULL);
171 env = cl_env_get(&refcheck);
173 RETURN(PTR_ERR(env));
175 io = ll_fault_io_init(env, vma, vmpage->index, NULL);
177 GOTO(out, result = PTR_ERR(io));
179 result = io->ci_result;
181 GOTO(out_io, result);
183 io->u.ci_fault.ft_mkwrite = 1;
184 io->u.ci_fault.ft_writable = 1;
186 vio = vvp_env_io(env);
187 vio->u.fault.ft_vma = vma;
188 vio->u.fault.ft_vmpage = vmpage;
190 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
192 inode = vvp_object_inode(io->ci_obj);
193 lli = ll_i2info(inode);
195 result = cl_io_loop(env, io);
197 cfs_restore_sigs(set);
201 if (vmpage->mapping == NULL) {
204 /* page was truncated and lock was cancelled, return
205 * ENODATA so that VM_FAULT_NOPAGE will be returned
206 * to handle_mm_fault(). */
209 } else if (!PageDirty(vmpage)) {
210 /* race, the page has been cleaned by ptlrpcd after
211 * it was unlocked, it has to be added into dirty
212 * cache again otherwise this soon-to-dirty page won't
213 * consume any grants, even worse if this page is being
214 * transferred because it will break RPC checksum.
218 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
219 "been written out, retry.\n",
220 vmpage, vmpage->index);
227 ll_file_set_flag(lli, LLIF_DATA_MODIFIED);
234 cl_env_put(env, &refcheck);
235 CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
236 LASSERT(ergo(result == 0, PageLocked(vmpage)));
241 static inline int to_fault_error(int result)
245 result = VM_FAULT_LOCKED;
248 result = VM_FAULT_NOPAGE;
251 result = VM_FAULT_OOM;
254 result = VM_FAULT_SIGBUS;
261 * Lustre implementation of a vm_operations_struct::fault() method, called by
262 * VM to server page fault (both in kernel and user space).
264 * \param vma - is virtiual area struct related to page fault
265 * \param vmf - structure which describe type and address where hit fault
267 * \return allocated and filled _locked_ page for address
268 * \retval VM_FAULT_ERROR on general error
269 * \retval NOPAGE_OOM not have memory for allocate new page
271 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
275 struct vvp_io *vio = NULL;
277 unsigned long ra_flags;
283 env = cl_env_get(&refcheck);
285 RETURN(PTR_ERR(env));
287 if (ll_sbi_has_fast_read(ll_i2sbi(file_inode(vma->vm_file)))) {
289 ll_cl_add(vma->vm_file, env, NULL, LCC_MMAP);
290 fault_ret = filemap_fault(vma, vmf);
291 ll_cl_remove(vma->vm_file, env);
293 /* - If there is no error, then the page was found in cache and
295 * - If VM_FAULT_RETRY is set, the page existed but failed to
296 * lock. It will return to kernel and retry;
297 * - Otherwise, it should try normal fault under DLM lock. */
298 if ((fault_ret & VM_FAULT_RETRY) ||
299 !(fault_ret & VM_FAULT_ERROR))
300 GOTO(out, result = 0);
305 io = ll_fault_io_init(env, vma, vmf->pgoff, &ra_flags);
307 GOTO(out, result = PTR_ERR(io));
309 result = io->ci_result;
311 vio = vvp_env_io(env);
312 vio->u.fault.ft_vma = vma;
313 vio->u.fault.ft_vmpage = NULL;
314 vio->u.fault.ft_vmf = vmf;
315 vio->u.fault.ft_flags = 0;
316 vio->u.fault.ft_flags_valid = 0;
318 /* May call ll_readpage() */
319 ll_cl_add(vma->vm_file, env, io, LCC_MMAP);
321 result = cl_io_loop(env, io);
323 ll_cl_remove(vma->vm_file, env);
325 /* ft_flags are only valid if we reached
326 * the call to filemap_fault */
327 if (vio->u.fault.ft_flags_valid)
328 fault_ret = vio->u.fault.ft_flags;
330 vmpage = vio->u.fault.ft_vmpage;
331 if (result != 0 && vmpage != NULL) {
332 page_cache_release(vmpage);
338 vma->vm_flags |= ra_flags;
341 cl_env_put(env, &refcheck);
342 if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
343 fault_ret |= to_fault_error(result);
345 CDEBUG(D_MMAP, "%s fault %d/%d\n", current->comm, fault_ret, result);
349 static int ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
352 bool printed = false;
356 /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
357 * so that it can be killed by admin but not cause segfault by
359 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
361 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
365 result = ll_fault0(vma, vmf);
366 if (!(result & (VM_FAULT_RETRY | VM_FAULT_ERROR | VM_FAULT_LOCKED))) {
367 struct page *vmpage = vmf->page;
369 /* check if this page has been truncated */
371 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
373 page_cache_release(vmpage);
376 if (!printed && ++count > 16) {
377 CWARN("the page is under heavy contention,"
378 "maybe your app(%s) needs revising :-)\n",
386 result |= VM_FAULT_LOCKED;
388 cfs_restore_sigs(set);
392 static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
395 bool printed = false;
399 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
400 LPROC_LL_MKWRITE, 1);
402 file_update_time(vma->vm_file);
405 result = ll_page_mkwrite0(vma, vmf->page, &retry);
407 if (!printed && ++count > 16) {
408 const struct dentry *de = vma->vm_file->f_path.dentry;
410 CWARN("app(%s): the page %lu of file "DFID" is under"
411 " heavy contention\n",
412 current->comm, vmf->pgoff,
413 PFID(ll_inode2fid(de->d_inode)));
420 LASSERT(PageLocked(vmf->page));
421 result = VM_FAULT_LOCKED;
425 result = VM_FAULT_NOPAGE;
428 result = VM_FAULT_OOM;
431 result = VM_FAULT_RETRY;
434 result = VM_FAULT_SIGBUS;
442 * To avoid cancel the locks covering mmapped region for lock cache pressure,
443 * we track the mapped vma count in vvp_object::vob_mmap_cnt.
445 static void ll_vm_open(struct vm_area_struct * vma)
447 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
448 struct vvp_object *vob = cl_inode2vvp(inode);
451 LASSERT(vma->vm_file);
452 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
453 atomic_inc(&vob->vob_mmap_cnt);
458 * Dual to ll_vm_open().
460 static void ll_vm_close(struct vm_area_struct *vma)
462 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
463 struct vvp_object *vob = cl_inode2vvp(inode);
466 LASSERT(vma->vm_file);
467 atomic_dec(&vob->vob_mmap_cnt);
468 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
472 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
473 * nopage's reference passing to the pte */
474 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
479 LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
480 if (mapping_mapped(mapping)) {
482 unmap_mapping_range(mapping, first + PAGE_CACHE_SIZE - 1,
483 last - first + 1, 0);
489 static const struct vm_operations_struct ll_file_vm_ops = {
491 .page_mkwrite = ll_page_mkwrite,
493 .close = ll_vm_close,
496 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
498 struct inode *inode = file->f_path.dentry->d_inode;
502 if (ll_file_nolock(file))
505 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
506 rc = generic_file_mmap(file, vma);
508 vma->vm_ops = &ll_file_vm_ops;
509 vma->vm_ops->open(vma);
510 /* update the inode's size and mtime */
511 rc = ll_glimpse_size(inode);