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, 2014, 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(ldlm_policy_data_t *policy,
60 struct vm_area_struct *vma, unsigned long addr,
63 policy->l_extent.start = ((addr - vma->vm_start) & PAGE_MASK) +
64 (vma->vm_pgoff << PAGE_CACHE_SHIFT);
65 policy->l_extent.end = (policy->l_extent.start + count - 1) |
69 struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
72 struct vm_area_struct *vma, *ret = NULL;
75 /* mmap_sem must have been held by caller. */
76 LASSERT(!down_write_trylock(&mm->mmap_sem));
78 for(vma = find_vma(mm, addr);
79 vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
80 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
81 vma->vm_flags & VM_SHARED) {
90 * API independent part for page fault initialization.
91 * \param vma - virtual memory area addressed to page fault
92 * \param env - corespondent lu_env to processing
93 * \param nest - nested level
94 * \param index - page index corespondent to fault.
95 * \parm ra_flags - vma readahead flags.
97 * \return allocated and initialized env for fault operation.
98 * \retval EINVAL if env can't allocated
99 * \return other error codes from cl_io_init.
101 static struct cl_io *
102 ll_fault_io_init(struct vm_area_struct *vma, struct lu_env **env_ret,
103 struct cl_env_nest *nest, pgoff_t index,
104 unsigned long *ra_flags)
106 struct file *file = vma->vm_file;
107 struct inode *inode = file->f_path.dentry->d_inode;
109 struct cl_fault_io *fio;
115 if (ll_file_nolock(file))
116 RETURN(ERR_PTR(-EOPNOTSUPP));
119 * page fault can be called when lustre IO is
120 * already active for the current thread, e.g., when doing read/write
121 * against user level buffer mapped from Lustre buffer. To avoid
122 * stomping on existing context, optionally force an allocation of a new
125 env = cl_env_nested_get(nest);
127 RETURN(ERR_PTR(-EINVAL));
131 io = vvp_env_thread_io(env);
132 io->ci_obj = ll_i2info(inode)->lli_clob;
133 LASSERT(io->ci_obj != NULL);
135 fio = &io->u.ci_fault;
136 fio->ft_index = index;
137 fio->ft_executable = vma->vm_flags&VM_EXEC;
140 * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
141 * the kernel will not read other pages not covered by ldlm in
142 * filemap_nopage. we do our readahead in ll_readpage.
144 if (ra_flags != NULL)
145 *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
146 vma->vm_flags &= ~VM_SEQ_READ;
147 vma->vm_flags |= VM_RAND_READ;
149 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
150 fio->ft_index, fio->ft_executable);
152 rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
154 struct vvp_io *vio = vvp_env_io(env);
155 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
157 LASSERT(vio->vui_cl.cis_io == io);
159 /* mmap lock must be MANDATORY it has to cache
161 io->ci_lockreq = CILR_MANDATORY;
166 cl_env_nested_put(nest, env);
173 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
174 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
180 struct cl_env_nest nest;
184 struct ll_inode_info *lli;
187 LASSERT(vmpage != NULL);
189 io = ll_fault_io_init(vma, &env, &nest, vmpage->index, NULL);
191 GOTO(out, result = PTR_ERR(io));
193 result = io->ci_result;
195 GOTO(out_io, result);
197 io->u.ci_fault.ft_mkwrite = 1;
198 io->u.ci_fault.ft_writable = 1;
200 vio = vvp_env_io(env);
201 vio->u.fault.ft_vma = vma;
202 vio->u.fault.ft_vmpage = vmpage;
204 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
206 /* we grab lli_trunc_sem to exclude truncate case.
207 * Otherwise, we could add dirty pages into osc cache
208 * while truncate is on-going. */
209 inode = vvp_object_inode(io->ci_obj);
210 lli = ll_i2info(inode);
211 down_read(&lli->lli_trunc_sem);
213 result = cl_io_loop(env, io);
215 up_read(&lli->lli_trunc_sem);
217 cfs_restore_sigs(set);
221 if (vmpage->mapping == NULL) {
224 /* page was truncated and lock was cancelled, return
225 * ENODATA so that VM_FAULT_NOPAGE will be returned
226 * to handle_mm_fault(). */
229 } else if (!PageDirty(vmpage)) {
230 /* race, the page has been cleaned by ptlrpcd after
231 * it was unlocked, it has to be added into dirty
232 * cache again otherwise this soon-to-dirty page won't
233 * consume any grants, even worse if this page is being
234 * transferred because it will break RPC checksum.
238 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
239 "been written out, retry.\n",
240 vmpage, vmpage->index);
247 spin_lock(&lli->lli_lock);
248 lli->lli_flags |= LLIF_DATA_MODIFIED;
249 spin_unlock(&lli->lli_lock);
256 cl_env_nested_put(&nest, env);
258 CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
259 LASSERT(ergo(result == 0, PageLocked(vmpage)));
264 static inline int to_fault_error(int result)
268 result = VM_FAULT_LOCKED;
271 result = VM_FAULT_NOPAGE;
274 result = VM_FAULT_OOM;
277 result = VM_FAULT_SIGBUS;
284 * Lustre implementation of a vm_operations_struct::fault() method, called by
285 * VM to server page fault (both in kernel and user space).
287 * \param vma - is virtiual area struct related to page fault
288 * \param vmf - structure which describe type and address where hit fault
290 * \return allocated and filled _locked_ page for address
291 * \retval VM_FAULT_ERROR on general error
292 * \retval NOPAGE_OOM not have memory for allocate new page
294 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
298 struct vvp_io *vio = NULL;
300 unsigned long ra_flags;
301 struct cl_env_nest nest;
306 io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags);
308 RETURN(to_fault_error(PTR_ERR(io)));
310 result = io->ci_result;
312 vio = vvp_env_io(env);
313 vio->u.fault.ft_vma = vma;
314 vio->u.fault.ft_vmpage = NULL;
315 vio->u.fault.ft_vmf = vmf;
316 vio->u.fault.ft_flags = 0;
317 vio->u.fault.ft_flags_valid = 0;
319 /* May call ll_readpage() */
320 ll_cl_add(vma->vm_file, env, io);
322 result = cl_io_loop(env, io);
324 ll_cl_remove(vma->vm_file, env);
326 /* ft_flags are only valid if we reached
327 * the call to filemap_fault */
328 if (vio->u.fault.ft_flags_valid)
329 fault_ret = vio->u.fault.ft_flags;
331 vmpage = vio->u.fault.ft_vmpage;
332 if (result != 0 && vmpage != NULL) {
333 page_cache_release(vmpage);
338 cl_env_nested_put(&nest, env);
340 vma->vm_flags |= ra_flags;
341 if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
342 fault_ret |= to_fault_error(result);
344 CDEBUG(D_MMAP, "%s fault %d/%d\n",
345 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));
362 result = ll_fault0(vma, vmf);
363 LASSERT(!(result & VM_FAULT_LOCKED));
365 struct page *vmpage = vmf->page;
367 /* check if this page has been truncated */
369 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
371 page_cache_release(vmpage);
374 if (!printed && ++count > 16) {
375 CWARN("the page is under heavy contention,"
376 "maybe your app(%s) needs revising :-)\n",
384 result |= VM_FAULT_LOCKED;
386 cfs_restore_sigs(set);
390 static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
393 bool printed = false;
399 result = ll_page_mkwrite0(vma, vmf->page, &retry);
401 if (!printed && ++count > 16) {
402 const struct dentry *de = vma->vm_file->f_path.dentry;
404 CWARN("app(%s): the page %lu of file "DFID" is under"
405 " heavy contention\n",
406 current->comm, vmf->pgoff,
407 PFID(ll_inode2fid(de->d_inode)));
414 LASSERT(PageLocked(vmf->page));
415 result = VM_FAULT_LOCKED;
419 result = VM_FAULT_NOPAGE;
422 result = VM_FAULT_OOM;
425 result = VM_FAULT_RETRY;
428 result = VM_FAULT_SIGBUS;
436 * To avoid cancel the locks covering mmapped region for lock cache pressure,
437 * we track the mapped vma count in vvp_object::vob_mmap_cnt.
439 static void ll_vm_open(struct vm_area_struct * vma)
441 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
442 struct vvp_object *vob = cl_inode2vvp(inode);
445 LASSERT(vma->vm_file);
446 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
447 atomic_inc(&vob->vob_mmap_cnt);
452 * Dual to ll_vm_open().
454 static void ll_vm_close(struct vm_area_struct *vma)
456 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
457 struct vvp_object *vob = cl_inode2vvp(inode);
460 LASSERT(vma->vm_file);
461 atomic_dec(&vob->vob_mmap_cnt);
462 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
466 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
467 * nopage's reference passing to the pte */
468 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
473 LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
474 if (mapping_mapped(mapping)) {
476 unmap_mapping_range(mapping, first + PAGE_CACHE_SIZE - 1,
477 last - first + 1, 0);
483 static const struct vm_operations_struct ll_file_vm_ops = {
485 .page_mkwrite = ll_page_mkwrite,
487 .close = ll_vm_close,
490 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
492 struct inode *inode = file->f_path.dentry->d_inode;
496 if (ll_file_nolock(file))
499 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
500 rc = generic_file_mmap(file, vma);
502 vma->vm_ops = &ll_file_vm_ops;
503 vma->vm_ops->open(vma);
504 /* update the inode's size and mtime */
505 rc = ll_glimpse_size(inode);