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 vma - virtual memory area addressed to page fault
91 * \param env - corespondent lu_env to processing
92 * \param nest - nested level
93 * \param index - page index corespondent to fault.
94 * \parm ra_flags - vma readahead flags.
96 * \return allocated and initialized env for fault operation.
97 * \retval EINVAL if env can't allocated
98 * \return other error codes from cl_io_init.
100 static struct cl_io *
101 ll_fault_io_init(struct vm_area_struct *vma, struct lu_env **env_ret,
102 struct cl_env_nest *nest, pgoff_t index,
103 unsigned long *ra_flags)
105 struct file *file = vma->vm_file;
106 struct inode *inode = file->f_path.dentry->d_inode;
108 struct cl_fault_io *fio;
114 if (ll_file_nolock(file))
115 RETURN(ERR_PTR(-EOPNOTSUPP));
118 * page fault can be called when lustre IO is
119 * already active for the current thread, e.g., when doing read/write
120 * against user level buffer mapped from Lustre buffer. To avoid
121 * stomping on existing context, optionally force an allocation of a new
124 env = cl_env_nested_get(nest);
126 RETURN(ERR_PTR(-EINVAL));
130 io = vvp_env_thread_io(env);
131 io->ci_obj = ll_i2info(inode)->lli_clob;
132 LASSERT(io->ci_obj != NULL);
134 fio = &io->u.ci_fault;
135 fio->ft_index = index;
136 fio->ft_executable = vma->vm_flags&VM_EXEC;
139 * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
140 * the kernel will not read other pages not covered by ldlm in
141 * filemap_nopage. we do our readahead in ll_readpage.
143 if (ra_flags != NULL)
144 *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
145 vma->vm_flags &= ~VM_SEQ_READ;
146 vma->vm_flags |= VM_RAND_READ;
148 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
149 fio->ft_index, fio->ft_executable);
151 rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
153 struct vvp_io *vio = vvp_env_io(env);
154 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
156 LASSERT(vio->vui_cl.cis_io == io);
158 /* mmap lock must be MANDATORY it has to cache
160 io->ci_lockreq = CILR_MANDATORY;
165 cl_env_nested_put(nest, env);
172 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
173 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
179 struct cl_env_nest nest;
183 struct ll_inode_info *lli;
186 LASSERT(vmpage != NULL);
188 io = ll_fault_io_init(vma, &env, &nest, vmpage->index, NULL);
190 GOTO(out, result = PTR_ERR(io));
192 result = io->ci_result;
194 GOTO(out_io, result);
196 io->u.ci_fault.ft_mkwrite = 1;
197 io->u.ci_fault.ft_writable = 1;
199 vio = vvp_env_io(env);
200 vio->u.fault.ft_vma = vma;
201 vio->u.fault.ft_vmpage = vmpage;
203 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
205 /* we grab lli_trunc_sem to exclude truncate case.
206 * Otherwise, we could add dirty pages into osc cache
207 * while truncate is on-going. */
208 inode = vvp_object_inode(io->ci_obj);
209 lli = ll_i2info(inode);
210 down_read(&lli->lli_trunc_sem);
212 result = cl_io_loop(env, io);
214 up_read(&lli->lli_trunc_sem);
216 cfs_restore_sigs(set);
220 if (vmpage->mapping == NULL) {
223 /* page was truncated and lock was cancelled, return
224 * ENODATA so that VM_FAULT_NOPAGE will be returned
225 * to handle_mm_fault(). */
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.
237 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
238 "been written out, retry.\n",
239 vmpage, vmpage->index);
246 ll_file_set_flag(lli, LLIF_DATA_MODIFIED);
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)));
260 static inline int to_fault_error(int result)
264 result = VM_FAULT_LOCKED;
267 result = VM_FAULT_NOPAGE;
270 result = VM_FAULT_OOM;
273 result = VM_FAULT_SIGBUS;
280 * Lustre implementation of a vm_operations_struct::fault() method, called by
281 * VM to server page fault (both in kernel and user space).
283 * \param vma - is virtiual area struct related to page fault
284 * \param vmf - structure which describe type and address where hit fault
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
290 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
294 struct vvp_io *vio = NULL;
296 unsigned long ra_flags;
297 struct cl_env_nest nest;
302 io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags);
304 RETURN(to_fault_error(PTR_ERR(io)));
306 result = io->ci_result;
308 vio = vvp_env_io(env);
309 vio->u.fault.ft_vma = vma;
310 vio->u.fault.ft_vmpage = NULL;
311 vio->u.fault.ft_vmf = vmf;
312 vio->u.fault.ft_flags = 0;
313 vio->u.fault.ft_flags_valid = 0;
315 /* May call ll_readpage() */
316 ll_cl_add(vma->vm_file, env, io);
318 result = cl_io_loop(env, io);
320 ll_cl_remove(vma->vm_file, env);
322 /* ft_flags are only valid if we reached
323 * the call to filemap_fault */
324 if (vio->u.fault.ft_flags_valid)
325 fault_ret = vio->u.fault.ft_flags;
327 vmpage = vio->u.fault.ft_vmpage;
328 if (result != 0 && vmpage != NULL) {
329 page_cache_release(vmpage);
334 cl_env_nested_put(&nest, env);
336 vma->vm_flags |= ra_flags;
337 if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
338 fault_ret |= to_fault_error(result);
340 CDEBUG(D_MMAP, "%s fault %d/%d\n",
341 current->comm, fault_ret, result);
345 static int ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
348 bool printed = false;
352 /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
353 * so that it can be killed by admin but not cause segfault by
355 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
358 result = ll_fault0(vma, vmf);
359 LASSERT(!(result & VM_FAULT_LOCKED));
361 struct page *vmpage = vmf->page;
363 /* check if this page has been truncated */
365 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
367 page_cache_release(vmpage);
370 if (!printed && ++count > 16) {
371 CWARN("the page is under heavy contention,"
372 "maybe your app(%s) needs revising :-)\n",
380 result |= VM_FAULT_LOCKED;
382 cfs_restore_sigs(set);
386 static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
389 bool printed = false;
395 result = ll_page_mkwrite0(vma, vmf->page, &retry);
397 if (!printed && ++count > 16) {
398 const struct dentry *de = vma->vm_file->f_path.dentry;
400 CWARN("app(%s): the page %lu of file "DFID" is under"
401 " heavy contention\n",
402 current->comm, vmf->pgoff,
403 PFID(ll_inode2fid(de->d_inode)));
410 LASSERT(PageLocked(vmf->page));
411 result = VM_FAULT_LOCKED;
415 result = VM_FAULT_NOPAGE;
418 result = VM_FAULT_OOM;
421 result = VM_FAULT_RETRY;
424 result = VM_FAULT_SIGBUS;
432 * To avoid cancel the locks covering mmapped region for lock cache pressure,
433 * we track the mapped vma count in vvp_object::vob_mmap_cnt.
435 static void ll_vm_open(struct vm_area_struct * vma)
437 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
438 struct vvp_object *vob = cl_inode2vvp(inode);
441 LASSERT(vma->vm_file);
442 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
443 atomic_inc(&vob->vob_mmap_cnt);
448 * Dual to ll_vm_open().
450 static void ll_vm_close(struct vm_area_struct *vma)
452 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
453 struct vvp_object *vob = cl_inode2vvp(inode);
456 LASSERT(vma->vm_file);
457 atomic_dec(&vob->vob_mmap_cnt);
458 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
462 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
463 * nopage's reference passing to the pte */
464 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
469 LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
470 if (mapping_mapped(mapping)) {
472 unmap_mapping_range(mapping, first + PAGE_CACHE_SIZE - 1,
473 last - first + 1, 0);
479 static const struct vm_operations_struct ll_file_vm_ops = {
481 .page_mkwrite = ll_page_mkwrite,
483 .close = ll_vm_close,
486 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
488 struct inode *inode = file->f_path.dentry->d_inode;
492 if (ll_file_nolock(file))
495 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
496 rc = generic_file_mmap(file, vma);
498 vma->vm_ops = &ll_file_vm_ops;
499 vma->vm_ops->open(vma);
500 /* update the inode's size and mtime */
501 rc = ll_glimpse_size(inode);