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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #include <linux/kernel.h>
35 #include <linux/string.h>
36 #include <linux/stat.h>
37 #include <linux/errno.h>
38 #include <linux/unistd.h>
39 #include <linux/uaccess.h>
42 #include <linux/pagemap.h>
44 #define DEBUG_SUBSYSTEM S_LLITE
46 #include "llite_internal.h"
48 static const struct vm_operations_struct ll_file_vm_ops;
50 void policy_from_vma(union ldlm_policy_data *policy,
51 struct vm_area_struct *vma, unsigned long addr,
54 policy->l_extent.start = ((addr - vma->vm_start) & PAGE_MASK) +
55 (vma->vm_pgoff << PAGE_SHIFT);
56 policy->l_extent.end = (policy->l_extent.start + count - 1) |
60 struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
63 struct vm_area_struct *vma, *ret = NULL;
65 /* mmap_sem must have been held by caller. */
66 LASSERT(!down_write_trylock(&mm->mmap_sem));
68 for (vma = find_vma(mm, addr);
69 vma && vma->vm_start < (addr + count); vma = vma->vm_next) {
70 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
71 vma->vm_flags & VM_SHARED) {
80 * API independent part for page fault initialization.
81 * \param vma - virtual memory area addressed to page fault
82 * \param env - corespondent lu_env to processing
83 * \param index - page index corespondent to fault.
84 * \parm ra_flags - vma readahead flags.
86 * \return error codes from cl_io_init.
89 ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma,
90 pgoff_t index, unsigned long *ra_flags)
92 struct file *file = vma->vm_file;
93 struct inode *inode = file_inode(file);
95 struct cl_fault_io *fio;
98 if (ll_file_nolock(file))
99 return ERR_PTR(-EOPNOTSUPP);
102 io = vvp_env_thread_io(env);
103 io->ci_obj = ll_i2info(inode)->lli_clob;
106 fio = &io->u.ci_fault;
107 fio->ft_index = index;
108 fio->ft_executable = vma->vm_flags & VM_EXEC;
111 * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
112 * the kernel will not read other pages not covered by ldlm in
113 * filemap_nopage. we do our readahead in ll_readpage.
116 *ra_flags = vma->vm_flags & (VM_RAND_READ | VM_SEQ_READ);
117 vma->vm_flags &= ~VM_SEQ_READ;
118 vma->vm_flags |= VM_RAND_READ;
120 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
121 fio->ft_index, fio->ft_executable);
123 rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
125 struct vvp_io *vio = vvp_env_io(env);
126 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
128 LASSERT(vio->vui_cl.cis_io == io);
130 /* mmap lock must be MANDATORY it has to cache pages. */
131 io->ci_lockreq = CILR_MANDATORY;
136 if (io->ci_need_restart)
145 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
146 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
156 struct ll_inode_info *lli;
158 env = cl_env_get(&refcheck);
162 io = ll_fault_io_init(env, vma, vmpage->index, NULL);
164 result = PTR_ERR(io);
168 result = io->ci_result;
172 io->u.ci_fault.ft_mkwrite = 1;
173 io->u.ci_fault.ft_writable = 1;
175 vio = vvp_env_io(env);
176 vio->u.fault.ft_vma = vma;
177 vio->u.fault.ft_vmpage = vmpage;
179 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
181 inode = vvp_object_inode(io->ci_obj);
182 lli = ll_i2info(inode);
184 result = cl_io_loop(env, io);
186 cfs_restore_sigs(set);
189 struct inode *inode = file_inode(vma->vm_file);
190 struct ll_inode_info *lli = ll_i2info(inode);
193 if (!vmpage->mapping) {
196 /* page was truncated and lock was cancelled, return
197 * ENODATA so that VM_FAULT_NOPAGE will be returned
198 * to handle_mm_fault().
202 } else if (!PageDirty(vmpage)) {
203 /* race, the page has been cleaned by ptlrpcd after
204 * it was unlocked, it has to be added into dirty
205 * cache again otherwise this soon-to-dirty page won't
206 * consume any grants, even worse if this page is being
207 * transferred because it will break RPC checksum.
211 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has been written out, retry.\n",
212 vmpage, vmpage->index);
219 set_bit(LLIF_DATA_MODIFIED, &lli->lli_flags);
225 cl_env_put(env, &refcheck);
226 CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
227 LASSERT(ergo(result == 0, PageLocked(vmpage)));
232 static inline int to_fault_error(int result)
236 result = VM_FAULT_LOCKED;
239 result = VM_FAULT_NOPAGE;
242 result = VM_FAULT_OOM;
245 result = VM_FAULT_SIGBUS;
252 * Lustre implementation of a vm_operations_struct::fault() method, called by
253 * VM to server page fault (both in kernel and user space).
255 * \param vma - is virtual area struct related to page fault
256 * \param vmf - structure which describe type and address where hit fault
258 * \return allocated and filled _locked_ page for address
259 * \retval VM_FAULT_ERROR on general error
260 * \retval NOPAGE_OOM not have memory for allocate new page
262 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
266 struct vvp_io *vio = NULL;
268 unsigned long ra_flags;
273 env = cl_env_get(&refcheck);
277 io = ll_fault_io_init(env, vma, vmf->pgoff, &ra_flags);
279 result = to_fault_error(PTR_ERR(io));
283 result = io->ci_result;
285 vio = vvp_env_io(env);
286 vio->u.fault.ft_vma = vma;
287 vio->u.fault.ft_vmpage = NULL;
288 vio->u.fault.ft_vmf = vmf;
289 vio->u.fault.ft_flags = 0;
290 vio->u.fault.ft_flags_valid = false;
292 /* May call ll_readpage() */
293 ll_cl_add(vma->vm_file, env, io);
295 result = cl_io_loop(env, io);
297 ll_cl_remove(vma->vm_file, env);
299 /* ft_flags are only valid if we reached
300 * the call to filemap_fault
302 if (vio->u.fault.ft_flags_valid)
303 fault_ret = vio->u.fault.ft_flags;
305 vmpage = vio->u.fault.ft_vmpage;
306 if (result != 0 && vmpage) {
313 vma->vm_flags |= ra_flags;
316 cl_env_put(env, &refcheck);
317 if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
318 fault_ret |= to_fault_error(result);
320 CDEBUG(D_MMAP, "%s fault %d/%d\n", current->comm, fault_ret, result);
324 static int ll_fault(struct vm_fault *vmf)
327 bool printed = false;
331 /* Only SIGKILL and SIGTERM are allowed for fault/nopage/mkwrite
332 * so that it can be killed by admin but not cause segfault by
335 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
338 result = ll_fault0(vmf->vma, vmf);
339 LASSERT(!(result & VM_FAULT_LOCKED));
341 struct page *vmpage = vmf->page;
343 /* check if this page has been truncated */
345 if (unlikely(!vmpage->mapping)) { /* unlucky */
350 if (!printed && ++count > 16) {
351 CWARN("the page is under heavy contention, maybe your app(%s) needs revising :-)\n",
359 result = VM_FAULT_LOCKED;
361 cfs_restore_sigs(set);
365 static int ll_page_mkwrite(struct vm_fault *vmf)
367 struct vm_area_struct *vma = vmf->vma;
369 bool printed = false;
373 file_update_time(vma->vm_file);
376 result = ll_page_mkwrite0(vma, vmf->page, &retry);
378 if (!printed && ++count > 16) {
379 const struct dentry *de = vma->vm_file->f_path.dentry;
381 CWARN("app(%s): the page %lu of file "DFID" is under heavy contention\n",
382 current->comm, vmf->pgoff,
383 PFID(ll_inode2fid(de->d_inode)));
390 LASSERT(PageLocked(vmf->page));
391 result = VM_FAULT_LOCKED;
396 result = VM_FAULT_NOPAGE;
399 result = VM_FAULT_OOM;
402 result = VM_FAULT_SIGBUS;
410 * To avoid cancel the locks covering mmapped region for lock cache pressure,
411 * we track the mapped vma count in vvp_object::vob_mmap_cnt.
413 static void ll_vm_open(struct vm_area_struct *vma)
415 struct inode *inode = file_inode(vma->vm_file);
416 struct vvp_object *vob = cl_inode2vvp(inode);
418 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
419 atomic_inc(&vob->vob_mmap_cnt);
423 * Dual to ll_vm_open().
425 static void ll_vm_close(struct vm_area_struct *vma)
427 struct inode *inode = file_inode(vma->vm_file);
428 struct vvp_object *vob = cl_inode2vvp(inode);
430 atomic_dec(&vob->vob_mmap_cnt);
431 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
434 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
435 * nopage's reference passing to the pte
437 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
441 LASSERTF(last > first, "last %llu first %llu\n", last, first);
442 if (mapping_mapped(mapping)) {
444 unmap_mapping_range(mapping, first + PAGE_SIZE - 1,
445 last - first + 1, 0);
451 static const struct vm_operations_struct ll_file_vm_ops = {
453 .page_mkwrite = ll_page_mkwrite,
455 .close = ll_vm_close,
458 int ll_file_mmap(struct file *file, struct vm_area_struct *vma)
460 struct inode *inode = file_inode(file);
463 if (ll_file_nolock(file))
466 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
467 rc = generic_file_mmap(file, vma);
469 vma->vm_ops = &ll_file_vm_ops;
470 vma->vm_ops->open(vma);
471 /* update the inode's size and mtime */
472 rc = ll_glimpse_size(inode);