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 <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 <lustre_lite.h>
55 #include "llite_internal.h"
56 #include <linux/lustre_compat25.h>
58 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
61 static struct vm_operations_struct ll_file_vm_ops;
63 void policy_from_vma(ldlm_policy_data_t *policy,
64 struct vm_area_struct *vma, unsigned long addr,
67 policy->l_extent.start = ((addr - vma->vm_start) & CFS_PAGE_MASK) +
68 (vma->vm_pgoff << PAGE_CACHE_SHIFT);
69 policy->l_extent.end = (policy->l_extent.start + count - 1) |
73 struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
76 struct vm_area_struct *vma, *ret = NULL;
79 /* mmap_sem must have been held by caller. */
80 LASSERT(!down_write_trylock(&mm->mmap_sem));
82 for(vma = find_vma(mm, addr);
83 vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
84 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
85 vma->vm_flags & VM_SHARED) {
94 * API independent part for page fault initialization.
95 * \param vma - virtual memory area addressed to page fault
96 * \param env - corespondent lu_env to processing
97 * \param nest - nested level
98 * \param index - page index corespondent to fault.
99 * \parm ra_flags - vma readahead flags.
101 * \return allocated and initialized env for fault operation.
102 * \retval EINVAL if env can't allocated
103 * \return other error codes from cl_io_init.
105 struct cl_io *ll_fault_io_init(struct vm_area_struct *vma,
106 struct lu_env **env_ret,
107 struct cl_env_nest *nest,
108 pgoff_t index, unsigned long *ra_flags)
110 struct file *file = vma->vm_file;
111 struct inode *inode = file->f_dentry->d_inode;
113 struct cl_fault_io *fio;
119 if (ll_file_nolock(file))
120 RETURN(ERR_PTR(-EOPNOTSUPP));
123 * page fault can be called when lustre IO is
124 * already active for the current thread, e.g., when doing read/write
125 * against user level buffer mapped from Lustre buffer. To avoid
126 * stomping on existing context, optionally force an allocation of a new
129 env = cl_env_nested_get(nest);
131 RETURN(ERR_PTR(-EINVAL));
135 io = ccc_env_thread_io(env);
136 io->ci_obj = ll_i2info(inode)->lli_clob;
137 LASSERT(io->ci_obj != NULL);
139 fio = &io->u.ci_fault;
140 fio->ft_index = index;
141 fio->ft_executable = vma->vm_flags&VM_EXEC;
144 * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
145 * the kernel will not read other pages not covered by ldlm in
146 * filemap_nopage. we do our readahead in ll_readpage.
148 if (ra_flags != NULL)
149 *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
150 vma->vm_flags &= ~VM_SEQ_READ;
151 vma->vm_flags |= VM_RAND_READ;
153 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
154 fio->ft_index, fio->ft_executable);
156 rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
158 struct ccc_io *cio = ccc_env_io(env);
159 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
161 LASSERT(cio->cui_cl.cis_io == io);
163 /* mmap lock must be MANDATORY it has to cache
165 io->ci_lockreq = CILR_MANDATORY;
170 cl_env_nested_put(nest, env);
177 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
178 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
184 struct cl_env_nest nest;
188 struct ll_inode_info *lli;
191 LASSERT(vmpage != NULL);
193 io = ll_fault_io_init(vma, &env, &nest, vmpage->index, NULL);
195 GOTO(out, result = PTR_ERR(io));
197 result = io->ci_result;
199 GOTO(out_io, result);
201 io->u.ci_fault.ft_mkwrite = 1;
202 io->u.ci_fault.ft_writable = 1;
204 vio = vvp_env_io(env);
205 vio->u.fault.ft_vma = vma;
206 vio->u.fault.ft_vmpage = vmpage;
208 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
210 /* we grab lli_trunc_sem to exclude truncate case.
211 * Otherwise, we could add dirty pages into osc cache
212 * while truncate is on-going. */
213 inode = ccc_object_inode(io->ci_obj);
214 lli = ll_i2info(inode);
215 down_read(&lli->lli_trunc_sem);
217 result = cl_io_loop(env, io);
219 up_read(&lli->lli_trunc_sem);
221 cfs_restore_sigs(set);
224 struct inode *inode = vma->vm_file->f_dentry->d_inode;
225 struct ll_inode_info *lli = ll_i2info(inode);
228 if (vmpage->mapping == NULL) {
231 /* page was truncated and lock was cancelled, return
232 * ENODATA so that VM_FAULT_NOPAGE will be returned
233 * to handle_mm_fault(). */
236 } else if (!PageDirty(vmpage)) {
237 /* race, the page has been cleaned by ptlrpcd after
238 * it was unlocked, it has to be added into dirty
239 * cache again otherwise this soon-to-dirty page won't
240 * consume any grants, even worse if this page is being
241 * transferred because it will break RPC checksum.
245 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
246 "been written out, retry.\n",
247 vmpage, vmpage->index);
254 spin_lock(&lli->lli_lock);
255 lli->lli_flags |= LLIF_DATA_MODIFIED;
256 spin_unlock(&lli->lli_lock);
263 cl_env_nested_put(&nest, env);
265 CDEBUG(D_MMAP, "%s mkwrite with %d\n", cfs_current()->comm, result);
266 LASSERT(ergo(result == 0, PageLocked(vmpage)));
272 #ifndef HAVE_VM_OP_FAULT
274 * Lustre implementation of a vm_operations_struct::nopage() method, called by
275 * VM to server page fault (both in kernel and user space).
277 * This function sets up CIT_FAULT cl_io that does the job.
279 * \param vma - is virtiual area struct related to page fault
280 * \param address - address when hit fault
281 * \param type - of fault
283 * \return allocated and filled _unlocked_ page for address
284 * \retval NOPAGE_SIGBUS if page not exist on this address
285 * \retval NOPAGE_OOM not have memory for allocate new page
287 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
291 struct cl_env_nest nest;
293 struct page *page = NOPAGE_SIGBUS;
294 struct vvp_io *vio = NULL;
295 unsigned long ra_flags;
298 const unsigned long writable = VM_SHARED|VM_WRITE;
302 pg_offset = ((address - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
303 io = ll_fault_io_init(vma, &env, &nest, pg_offset, &ra_flags);
305 return NOPAGE_SIGBUS;
307 result = io->ci_result;
311 io->u.ci_fault.ft_writable = (vma->vm_flags&writable) == writable;
313 vio = vvp_env_io(env);
314 vio->u.fault.ft_vma = vma;
315 vio->u.fault.nopage.ft_address = address;
316 vio->u.fault.nopage.ft_type = type;
317 vio->u.fault.ft_vmpage = NULL;
319 set = cfs_block_sigsinv(sigmask(SIGKILL)|sigmask(SIGTERM));
320 result = cl_io_loop(env, io);
321 cfs_restore_sigs(set);
323 page = vio->u.fault.ft_vmpage;
324 if (result != 0 && page != NULL) {
325 page_cache_release(page);
326 page = NOPAGE_SIGBUS;
330 if (result == -ENOMEM)
333 vma->vm_flags &= ~VM_RAND_READ;
334 vma->vm_flags |= ra_flags;
337 cl_env_nested_put(&nest, env);
344 static inline int to_fault_error(int result)
348 result = VM_FAULT_LOCKED;
351 result = VM_FAULT_NOPAGE;
354 result = VM_FAULT_OOM;
357 result = VM_FAULT_SIGBUS;
364 * Lustre implementation of a vm_operations_struct::fault() method, called by
365 * VM to server page fault (both in kernel and user space).
367 * \param vma - is virtiual area struct related to page fault
368 * \param vmf - structure which describe type and address where hit fault
370 * \return allocated and filled _locked_ page for address
371 * \retval VM_FAULT_ERROR on general error
372 * \retval NOPAGE_OOM not have memory for allocate new page
374 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
378 struct vvp_io *vio = NULL;
380 unsigned long ra_flags;
381 struct cl_env_nest nest;
386 io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags);
388 RETURN(to_fault_error(PTR_ERR(io)));
390 result = io->ci_result;
392 vio = vvp_env_io(env);
393 vio->u.fault.ft_vma = vma;
394 vio->u.fault.ft_vmpage = NULL;
395 vio->u.fault.fault.ft_vmf = vmf;
397 result = cl_io_loop(env, io);
399 fault_ret = vio->u.fault.fault.ft_flags;
400 vmpage = vio->u.fault.ft_vmpage;
401 if (result != 0 && vmpage != NULL) {
402 page_cache_release(vmpage);
407 cl_env_nested_put(&nest, env);
409 vma->vm_flags |= ra_flags;
410 if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
411 fault_ret |= to_fault_error(result);
413 CDEBUG(D_MMAP, "%s fault %d/%d\n",
414 cfs_current()->comm, fault_ret, result);
418 static int ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
421 bool printed = false;
425 /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
426 * so that it can be killed by admin but not cause segfault by
428 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
431 result = ll_fault0(vma, vmf);
432 LASSERT(!(result & VM_FAULT_LOCKED));
434 struct page *vmpage = vmf->page;
436 /* check if this page has been truncated */
438 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
440 page_cache_release(vmpage);
443 if (!printed && ++count > 16) {
444 CWARN("the page is under heavy contention,"
445 "maybe your app(%s) needs revising :-)\n",
453 result |= VM_FAULT_LOCKED;
455 cfs_restore_sigs(set);
460 #ifndef HAVE_PGMKWRITE_USE_VMFAULT
461 static int ll_page_mkwrite(struct vm_area_struct *vma, struct page *vmpage)
464 bool printed = false;
470 result = ll_page_mkwrite0(vma, vmpage, &retry);
472 if (!printed && ++count > 16) {
473 CWARN("app(%s): the page %lu of file %lu is under heavy"
475 current->comm, page_index(vmpage),
476 vma->vm_file->f_dentry->d_inode->i_ino);
483 else if (result == -ENODATA)
484 result = 0; /* kernel will know truncate has happened and
490 static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
493 bool printed = false;
499 result = ll_page_mkwrite0(vma, vmf->page, &retry);
501 if (!printed && ++count > 16) {
502 CWARN("app(%s): the page %lu of file %lu is under heavy"
504 current->comm, vmf->pgoff,
505 vma->vm_file->f_dentry->d_inode->i_ino);
512 LASSERT(PageLocked(vmf->page));
513 result = VM_FAULT_LOCKED;
517 result = VM_FAULT_NOPAGE;
520 result = VM_FAULT_OOM;
523 result = VM_FAULT_RETRY;
526 result = VM_FAULT_SIGBUS;
535 * To avoid cancel the locks covering mmapped region for lock cache pressure,
536 * we track the mapped vma count in ccc_object::cob_mmap_cnt.
538 static void ll_vm_open(struct vm_area_struct * vma)
540 struct inode *inode = vma->vm_file->f_dentry->d_inode;
541 struct ccc_object *vob = cl_inode2ccc(inode);
544 LASSERT(vma->vm_file);
545 LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
546 cfs_atomic_inc(&vob->cob_mmap_cnt);
551 * Dual to ll_vm_open().
553 static void ll_vm_close(struct vm_area_struct *vma)
555 struct inode *inode = vma->vm_file->f_dentry->d_inode;
556 struct ccc_object *vob = cl_inode2ccc(inode);
559 LASSERT(vma->vm_file);
560 cfs_atomic_dec(&vob->cob_mmap_cnt);
561 LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
565 #ifndef HAVE_VM_OP_FAULT
566 #ifndef HAVE_FILEMAP_POPULATE
567 static int (*filemap_populate)(struct vm_area_struct * area, unsigned long address, unsigned long len, pgprot_t prot, unsigned long pgoff, int nonblock);
569 static int ll_populate(struct vm_area_struct *area, unsigned long address,
570 unsigned long len, pgprot_t prot, unsigned long pgoff,
576 /* always set nonblock as true to avoid page read ahead */
577 rc = filemap_populate(area, address, len, prot, pgoff, 1);
582 /* return the user space pointer that maps to a file offset via a vma */
583 static inline unsigned long file_to_user(struct vm_area_struct *vma, __u64 byte)
585 return vma->vm_start +
586 (byte - ((__u64)vma->vm_pgoff << PAGE_CACHE_SHIFT));
590 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
591 * nopage's reference passing to the pte */
592 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
597 LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
598 if (mapping_mapped(mapping)) {
600 unmap_mapping_range(mapping, first + PAGE_CACHE_SIZE - 1,
601 last - first + 1, 0);
607 static struct vm_operations_struct ll_file_vm_ops = {
608 #ifndef HAVE_VM_OP_FAULT
610 .populate = ll_populate,
614 #ifndef HAVE_PGMKWRITE_COMPACT
615 .page_mkwrite = ll_page_mkwrite,
617 ._pmkw.page_mkwrite = ll_page_mkwrite,
620 .close = ll_vm_close,
623 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
625 struct inode *inode = file->f_dentry->d_inode;
629 if (ll_file_nolock(file))
632 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
633 rc = generic_file_mmap(file, vma);
635 #if !defined(HAVE_FILEMAP_POPULATE) && !defined(HAVE_VM_OP_FAULT)
636 if (!filemap_populate)
637 filemap_populate = vma->vm_ops->populate;
639 vma->vm_ops = &ll_file_vm_ops;
640 vma->vm_ops->open(vma);
641 /* update the inode's size and mtime */
642 rc = ll_glimpse_size(inode);