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, Whamcloud, Inc.
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 << CFS_PAGE_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;
118 if (ll_file_nolock(file))
119 RETURN(ERR_PTR(-EOPNOTSUPP));
122 * page fault can be called when lustre IO is
123 * already active for the current thread, e.g., when doing read/write
124 * against user level buffer mapped from Lustre buffer. To avoid
125 * stomping on existing context, optionally force an allocation of a new
128 env = cl_env_nested_get(nest);
130 RETURN(ERR_PTR(-EINVAL));
134 io = ccc_env_thread_io(env);
135 io->ci_obj = ll_i2info(inode)->lli_clob;
136 LASSERT(io->ci_obj != NULL);
138 fio = &io->u.ci_fault;
139 fio->ft_index = index;
140 fio->ft_executable = vma->vm_flags&VM_EXEC;
143 * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
144 * the kernel will not read other pages not covered by ldlm in
145 * filemap_nopage. we do our readahead in ll_readpage.
147 if (ra_flags != NULL)
148 *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
149 vma->vm_flags &= ~VM_SEQ_READ;
150 vma->vm_flags |= VM_RAND_READ;
152 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
153 fio->ft_index, fio->ft_executable);
155 if (cl_io_init(env, io, CIT_FAULT, io->ci_obj) == 0) {
156 struct ccc_io *cio = ccc_env_io(env);
157 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
159 LASSERT(cio->cui_cl.cis_io == io);
161 /* mmap lock must be MANDATORY
162 * it has to cache pages. */
163 io->ci_lockreq = CILR_MANDATORY;
171 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
172 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
178 struct cl_env_nest nest;
182 struct ll_inode_info *lli;
185 LASSERT(vmpage != NULL);
187 io = ll_fault_io_init(vma, &env, &nest, vmpage->index, NULL);
189 GOTO(out, result = PTR_ERR(io));
191 result = io->ci_result;
195 /* Don't enqueue new locks for page_mkwrite().
196 * If the lock has been cancelled then page must have been
197 * truncated, in that case, kernel will handle it.
199 io->ci_lockreq = CILR_PEEK;
200 io->u.ci_fault.ft_mkwrite = 1;
201 io->u.ci_fault.ft_writable = 1;
203 vio = vvp_env_io(env);
204 vio->u.fault.ft_vma = vma;
205 vio->u.fault.ft_vmpage = vmpage;
207 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
209 /* we grab lli_trunc_sem to exclude truncate case.
210 * Otherwise, we could add dirty pages into osc cache
211 * while truncate is on-going. */
212 inode = ccc_object_inode(io->ci_obj);
213 lli = ll_i2info(inode);
214 down_read(&lli->lli_trunc_sem);
216 result = cl_io_loop(env, io);
218 up_read(&lli->lli_trunc_sem);
220 cfs_restore_sigs(set);
222 if (result == -ENODATA) /* peek failed, no lock caching. */
223 CDEBUG(D_MMAP, "race on page_mkwrite: %lx (%lu %p)\n",
224 vma->vm_flags, io->u.ci_fault.ft_index, vmpage);
226 if (result == 0 || result == -ENODATA) {
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 (result == -ENODATA) {
237 /* Invalidate it if the cl_lock is being revoked.
238 * This piece of code is definitely needed for RHEL5,
239 * otherwise, SIGBUS will be wrongly returned to
241 write_one_page(vmpage, 1);
243 if (vmpage->mapping != NULL) {
244 ll_invalidate_page(vmpage);
245 LASSERT(vmpage->mapping == NULL);
248 } else if (!PageDirty(vmpage)) {
249 /* race, the page has been cleaned by ptlrpcd after
250 * it was unlocked, it has to be added into dirty
251 * cache again otherwise this soon-to-dirty page won't
252 * consume any grants, even worse if this page is being
253 * transferred because it will break RPC checksum.
257 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
258 "been written out, retry.\n",
259 vmpage, vmpage->index);
269 cl_env_nested_put(&nest, env);
271 CDEBUG(D_MMAP, "%s mkwrite with %d\n", cfs_current()->comm, result);
273 LASSERT(ergo(result == 0, PageLocked(vmpage)));
278 #ifndef HAVE_VM_OP_FAULT
280 * Lustre implementation of a vm_operations_struct::nopage() method, called by
281 * VM to server page fault (both in kernel and user space).
283 * This function sets up CIT_FAULT cl_io that does the job.
285 * \param vma - is virtiual area struct related to page fault
286 * \param address - address when hit fault
287 * \param type - of fault
289 * \return allocated and filled _unlocked_ page for address
290 * \retval NOPAGE_SIGBUS if page not exist on this address
291 * \retval NOPAGE_OOM not have memory for allocate new page
293 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
297 struct cl_env_nest nest;
299 struct page *page = NOPAGE_SIGBUS;
300 struct vvp_io *vio = NULL;
301 unsigned long ra_flags;
304 const unsigned long writable = VM_SHARED|VM_WRITE;
308 pg_offset = ((address - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
309 io = ll_fault_io_init(vma, &env, &nest, pg_offset, &ra_flags);
311 return NOPAGE_SIGBUS;
313 result = io->ci_result;
317 io->u.ci_fault.ft_writable = (vma->vm_flags&writable) == writable;
319 vio = vvp_env_io(env);
320 vio->u.fault.ft_vma = vma;
321 vio->u.fault.nopage.ft_address = address;
322 vio->u.fault.nopage.ft_type = type;
323 vio->u.fault.ft_vmpage = NULL;
325 set = cfs_block_sigsinv(sigmask(SIGKILL)|sigmask(SIGTERM));
326 result = cl_io_loop(env, io);
327 cfs_restore_sigs(set);
329 page = vio->u.fault.ft_vmpage;
330 if (result != 0 && page != NULL) {
331 page_cache_release(page);
332 page = NOPAGE_SIGBUS;
336 if (result == -ENOMEM)
339 vma->vm_flags &= ~VM_RAND_READ;
340 vma->vm_flags |= ra_flags;
343 cl_env_nested_put(&nest, env);
350 static inline int to_fault_error(int result)
354 result = VM_FAULT_LOCKED;
357 result = VM_FAULT_NOPAGE;
360 result = VM_FAULT_OOM;
363 result = VM_FAULT_SIGBUS;
370 * Lustre implementation of a vm_operations_struct::fault() method, called by
371 * VM to server page fault (both in kernel and user space).
373 * \param vma - is virtiual area struct related to page fault
374 * \param vmf - structure which describe type and address where hit fault
376 * \return allocated and filled _locked_ page for address
377 * \retval VM_FAULT_ERROR on general error
378 * \retval NOPAGE_OOM not have memory for allocate new page
380 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
384 struct vvp_io *vio = NULL;
386 unsigned long ra_flags;
387 struct cl_env_nest nest;
392 io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags);
394 RETURN(to_fault_error(PTR_ERR(io)));
396 result = io->ci_result;
398 vio = vvp_env_io(env);
399 vio->u.fault.ft_vma = vma;
400 vio->u.fault.ft_vmpage = NULL;
401 vio->u.fault.fault.ft_vmf = vmf;
403 result = cl_io_loop(env, io);
405 fault_ret = vio->u.fault.fault.ft_flags;
406 vmpage = vio->u.fault.ft_vmpage;
407 if (result != 0 && vmpage != NULL) {
408 page_cache_release(vmpage);
413 cl_env_nested_put(&nest, env);
415 vma->vm_flags |= ra_flags;
416 if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
417 fault_ret |= to_fault_error(result);
419 CDEBUG(D_MMAP, "%s fault %d/%d\n",
420 cfs_current()->comm, fault_ret, result);
424 static int ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
427 bool printed = false;
431 /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
432 * so that it can be killed by admin but not cause segfault by
434 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
437 result = ll_fault0(vma, vmf);
438 LASSERT(!(result & VM_FAULT_LOCKED));
440 struct page *vmpage = vmf->page;
442 /* check if this page has been truncated */
444 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
446 page_cache_release(vmpage);
449 if (!printed && ++count > 16) {
450 CWARN("the page is under heavy contention,"
451 "maybe your app(%s) needs revising :-)\n",
459 result |= VM_FAULT_LOCKED;
461 cfs_restore_sigs(set);
466 #ifndef HAVE_PGMKWRITE_USE_VMFAULT
467 static int ll_page_mkwrite(struct vm_area_struct *vma, struct page *vmpage)
470 bool printed = false;
476 result = ll_page_mkwrite0(vma, vmpage, &retry);
478 if (!printed && ++count > 16) {
479 CWARN("app(%s): the page %lu of file %lu is under heavy"
481 current->comm, page_index(vmpage),
482 vma->vm_file->f_dentry->d_inode->i_ino);
489 else if (result == -ENODATA)
490 result = 0; /* kernel will know truncate has happened and
496 static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
499 bool printed = false;
505 result = ll_page_mkwrite0(vma, vmf->page, &retry);
507 if (!printed && ++count > 16) {
508 CWARN("app(%s): the page %lu of file %lu is under heavy"
510 current->comm, vmf->pgoff,
511 vma->vm_file->f_dentry->d_inode->i_ino);
518 LASSERT(PageLocked(vmf->page));
519 result = VM_FAULT_LOCKED;
523 result = VM_FAULT_NOPAGE;
526 result = VM_FAULT_OOM;
529 result = VM_FAULT_RETRY;
532 result = VM_FAULT_SIGBUS;
541 * To avoid cancel the locks covering mmapped region for lock cache pressure,
542 * we track the mapped vma count in ccc_object::cob_mmap_cnt.
544 static void ll_vm_open(struct vm_area_struct * vma)
546 struct inode *inode = vma->vm_file->f_dentry->d_inode;
547 struct ccc_object *vob = cl_inode2ccc(inode);
550 LASSERT(vma->vm_file);
551 LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
552 cfs_atomic_inc(&vob->cob_mmap_cnt);
557 * Dual to ll_vm_open().
559 static void ll_vm_close(struct vm_area_struct *vma)
561 struct inode *inode = vma->vm_file->f_dentry->d_inode;
562 struct ccc_object *vob = cl_inode2ccc(inode);
565 LASSERT(vma->vm_file);
566 cfs_atomic_dec(&vob->cob_mmap_cnt);
567 LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
571 #ifndef HAVE_VM_OP_FAULT
572 #ifndef HAVE_FILEMAP_POPULATE
573 static int (*filemap_populate)(struct vm_area_struct * area, unsigned long address, unsigned long len, pgprot_t prot, unsigned long pgoff, int nonblock);
575 static int ll_populate(struct vm_area_struct *area, unsigned long address,
576 unsigned long len, pgprot_t prot, unsigned long pgoff,
582 /* always set nonblock as true to avoid page read ahead */
583 rc = filemap_populate(area, address, len, prot, pgoff, 1);
588 /* return the user space pointer that maps to a file offset via a vma */
589 static inline unsigned long file_to_user(struct vm_area_struct *vma, __u64 byte)
591 return vma->vm_start + (byte - ((__u64)vma->vm_pgoff << CFS_PAGE_SHIFT));
595 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
596 * nopage's reference passing to the pte */
597 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
602 LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
603 if (mapping_mapped(mapping)) {
605 unmap_mapping_range(mapping, first + CFS_PAGE_SIZE - 1,
606 last - first + 1, 0);
612 static struct vm_operations_struct ll_file_vm_ops = {
613 #ifndef HAVE_VM_OP_FAULT
615 .populate = ll_populate,
619 #ifndef HAVE_PGMKWRITE_COMPACT
620 .page_mkwrite = ll_page_mkwrite,
622 ._pmkw.page_mkwrite = ll_page_mkwrite,
625 .close = ll_vm_close,
628 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
630 struct inode *inode = file->f_dentry->d_inode;
634 if (ll_file_nolock(file))
637 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
638 rc = generic_file_mmap(file, vma);
640 #if !defined(HAVE_FILEMAP_POPULATE) && !defined(HAVE_VM_OP_FAULT)
641 if (!filemap_populate)
642 filemap_populate = vma->vm_ops->populate;
644 vma->vm_ops = &ll_file_vm_ops;
645 vma->vm_ops->open(vma);
646 /* update the inode's size and mtime */
647 rc = ll_glimpse_size(inode);