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 #ifndef AUTOCONF_INCLUDED
38 #include <linux/config.h>
40 #include <linux/kernel.h>
42 #include <linux/string.h>
43 #include <linux/stat.h>
44 #include <linux/errno.h>
45 #include <linux/smp_lock.h>
46 #include <linux/unistd.h>
47 #include <linux/version.h>
48 #include <asm/system.h>
49 #include <asm/uaccess.h>
52 #include <linux/stat.h>
53 #include <asm/uaccess.h>
55 #include <linux/pagemap.h>
56 #include <linux/smp_lock.h>
58 #define DEBUG_SUBSYSTEM S_LLITE
60 #include <lustre_lite.h>
61 #include "llite_internal.h"
62 #include <linux/lustre_compat25.h>
64 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
67 static struct vm_operations_struct ll_file_vm_ops;
69 void policy_from_vma(ldlm_policy_data_t *policy,
70 struct vm_area_struct *vma, unsigned long addr,
73 policy->l_extent.start = ((addr - vma->vm_start) & CFS_PAGE_MASK) +
74 (vma->vm_pgoff << CFS_PAGE_SHIFT);
75 policy->l_extent.end = (policy->l_extent.start + count - 1) |
79 struct vm_area_struct * our_vma(unsigned long addr, size_t count)
81 struct mm_struct *mm = current->mm;
82 struct vm_area_struct *vma, *ret = NULL;
85 /* No MM (e.g. NFS)? No vmas too. */
89 spin_lock(&mm->page_table_lock);
90 for(vma = find_vma(mm, addr);
91 vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
92 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
93 vma->vm_flags & VM_SHARED) {
98 spin_unlock(&mm->page_table_lock);
103 * API independent part for page fault initialization.
104 * \param vma - virtual memory area addressed to page fault
105 * \param env - corespondent lu_env to processing
106 * \param nest - nested level
107 * \param index - page index corespondent to fault.
108 * \parm ra_flags - vma readahead flags.
110 * \return allocated and initialized env for fault operation.
111 * \retval EINVAL if env can't allocated
112 * \return other error codes from cl_io_init.
114 struct cl_io *ll_fault_io_init(struct vm_area_struct *vma,
115 struct lu_env **env_ret,
116 struct cl_env_nest *nest,
117 pgoff_t index, unsigned long *ra_flags)
119 struct file *file = vma->vm_file;
120 struct inode *inode = file->f_dentry->d_inode;
122 struct cl_fault_io *fio;
127 if (ll_file_nolock(file))
128 RETURN(ERR_PTR(-EOPNOTSUPP));
131 * page fault can be called when lustre IO is
132 * already active for the current thread, e.g., when doing read/write
133 * against user level buffer mapped from Lustre buffer. To avoid
134 * stomping on existing context, optionally force an allocation of a new
137 env = cl_env_nested_get(nest);
139 RETURN(ERR_PTR(-EINVAL));
143 io = ccc_env_thread_io(env);
144 io->ci_obj = ll_i2info(inode)->lli_clob;
145 LASSERT(io->ci_obj != NULL);
147 fio = &io->u.ci_fault;
148 fio->ft_index = index;
149 fio->ft_executable = vma->vm_flags&VM_EXEC;
152 * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
153 * the kernel will not read other pages not covered by ldlm in
154 * filemap_nopage. we do our readahead in ll_readpage.
156 if (ra_flags != NULL)
157 *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
158 vma->vm_flags &= ~VM_SEQ_READ;
159 vma->vm_flags |= VM_RAND_READ;
161 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
162 fio->ft_index, fio->ft_executable);
164 if (cl_io_init(env, io, CIT_FAULT, io->ci_obj) == 0) {
165 struct ccc_io *cio = ccc_env_io(env);
166 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
168 LASSERT(cio->cui_cl.cis_io == io);
170 /* mmap lock must be MANDATORY
171 * it has to cache pages. */
172 io->ci_lockreq = CILR_MANDATORY;
180 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
181 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
187 struct cl_env_nest nest;
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;
201 /* Don't enqueue new locks for page_mkwrite().
202 * If the lock has been cancelled then page must have been
203 * truncated, in that case, kernel will handle it.
205 io->ci_lockreq = CILR_PEEK;
206 io->u.ci_fault.ft_mkwrite = 1;
207 io->u.ci_fault.ft_writable = 1;
209 vio = vvp_env_io(env);
210 vio->u.fault.ft_vma = vma;
211 vio->u.fault.ft_vmpage = vmpage;
213 result = cl_io_loop(env, io);
215 if (result == -ENODATA) /* peek failed, no lock caching. */
216 CDEBUG(D_MMAP, "race on page_mkwrite: %lx (%lu %p)\n",
217 vma->vm_flags, io->u.ci_fault.ft_index, vmpage);
219 if (result == 0 || result == -ENODATA) {
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 (result == -ENODATA) {
230 /* Invalidate it if the cl_lock is being revoked.
231 * This piece of code is definitely needed for RHEL5,
232 * otherwise, SIGBUS will be wrongly returned to
234 write_one_page(vmpage, 1);
236 if (vmpage->mapping != NULL) {
237 ll_invalidate_page(vmpage);
238 LASSERT(vmpage->mapping == NULL);
241 } else if (!PageDirty(vmpage)) {
242 /* race, the page has been cleaned by ptlrpcd after
243 * it was unlocked, it has to be added into dirty
244 * cache again otherwise this soon-to-dirty page won't
245 * consume any grants, even worse if this page is being
246 * transferred because it will break RPC checksum.
250 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
251 "been written out, retry.\n",
252 vmpage, vmpage->index);
262 cl_env_nested_put(&nest, env);
264 CDEBUG(D_MMAP, "%s mkwrite with %d\n", cfs_current()->comm, result);
266 LASSERT(ergo(result == 0, PageLocked(vmpage)));
271 #ifndef HAVE_VM_OP_FAULT
273 * Lustre implementation of a vm_operations_struct::nopage() method, called by
274 * VM to server page fault (both in kernel and user space).
276 * This function sets up CIT_FAULT cl_io that does the job.
278 * \param vma - is virtiual area struct related to page fault
279 * \param address - address when hit fault
280 * \param type - of fault
282 * \return allocated and filled _unlocked_ page for address
283 * \retval NOPAGE_SIGBUS if page not exist on this address
284 * \retval NOPAGE_OOM not have memory for allocate new page
286 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
290 struct cl_env_nest nest;
292 struct page *page = NOPAGE_SIGBUS;
293 struct vvp_io *vio = NULL;
294 unsigned long ra_flags;
297 const unsigned long writable = VM_SHARED|VM_WRITE;
300 pg_offset = ((address - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
301 io = ll_fault_io_init(vma, &env, &nest, pg_offset, &ra_flags);
303 return NOPAGE_SIGBUS;
305 result = io->ci_result;
309 io->u.ci_fault.ft_writable = (vma->vm_flags&writable) == writable;
311 vio = vvp_env_io(env);
312 vio->u.fault.ft_vma = vma;
313 vio->u.fault.nopage.ft_address = address;
314 vio->u.fault.nopage.ft_type = type;
315 vio->u.fault.ft_vmpage = NULL;
317 result = cl_io_loop(env, io);
318 page = vio->u.fault.ft_vmpage;
319 if (result != 0 && page != NULL)
320 page_cache_release(page);
323 if (result == -ENOMEM)
326 vma->vm_flags &= ~VM_RAND_READ;
327 vma->vm_flags |= ra_flags;
330 cl_env_nested_put(&nest, env);
337 * Lustre implementation of a vm_operations_struct::fault() method, called by
338 * VM to server page fault (both in kernel and user space).
340 * \param vma - is virtiual area struct related to page fault
341 * \param vmf - structure which describe type and address where hit fault
343 * \return allocated and filled _locked_ page for address
344 * \retval VM_FAULT_ERROR on general error
345 * \retval NOPAGE_OOM not have memory for allocate new page
347 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
351 struct vvp_io *vio = NULL;
353 unsigned long ra_flags;
354 struct cl_env_nest nest;
359 io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags);
361 RETURN(VM_FAULT_ERROR);
363 result = io->ci_result;
367 vio = vvp_env_io(env);
368 vio->u.fault.ft_vma = vma;
369 vio->u.fault.ft_vmpage = NULL;
370 vio->u.fault.fault.ft_vmf = vmf;
372 result = cl_io_loop(env, io);
374 vmpage = vio->u.fault.ft_vmpage;
375 if (result != 0 && vmpage != NULL) {
376 page_cache_release(vmpage);
380 fault_ret = vio->u.fault.fault.ft_flags;
383 if (result != 0 && fault_ret == 0)
384 fault_ret = VM_FAULT_ERROR;
386 vma->vm_flags |= ra_flags;
389 cl_env_nested_put(&nest, env);
391 CDEBUG(D_MMAP, "%s fault %d/%d\n",
392 cfs_current()->comm, fault_ret, result);
396 static int ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
399 bool printed = false;
403 result = ll_fault0(vma, vmf);
404 LASSERT(!(result & VM_FAULT_LOCKED));
406 struct page *vmpage = vmf->page;
408 /* check if this page has been truncated */
410 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
412 page_cache_release(vmpage);
415 if (!printed && ++count > 16) {
416 CWARN("the page is under heavy contention,"
417 "maybe your app(%s) needs revising :-)\n",
425 result |= VM_FAULT_LOCKED;
431 #ifndef HAVE_PGMKWRITE_USE_VMFAULT
432 static int ll_page_mkwrite(struct vm_area_struct *vma, struct page *vmpage)
435 bool printed = false;
441 result = ll_page_mkwrite0(vma, vmpage, &retry);
443 if (!printed && ++count > 16) {
444 CWARN("app(%s): the page %lu of file %lu is under heavy"
446 current->comm, page_index(vmpage),
447 vma->vm_file->f_dentry->d_inode->i_ino);
454 else if (result == -ENODATA)
455 result = 0; /* kernel will know truncate has happened and
461 static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
464 bool printed = false;
470 result = ll_page_mkwrite0(vma, vmf->page, &retry);
472 if (!printed && ++count > 16) {
473 CWARN("app(%s): the page %lu of file %lu is under heavy"
475 current->comm, vmf->pgoff,
476 vma->vm_file->f_dentry->d_inode->i_ino);
483 LASSERT(PageLocked(vmf->page));
484 result = VM_FAULT_LOCKED;
488 result = VM_FAULT_NOPAGE;
491 result = VM_FAULT_OOM;
494 result = VM_FAULT_RETRY;
497 result = VM_FAULT_SIGBUS;
506 * To avoid cancel the locks covering mmapped region for lock cache pressure,
507 * we track the mapped vma count in ccc_object::cob_mmap_cnt.
509 static void ll_vm_open(struct vm_area_struct * vma)
511 struct inode *inode = vma->vm_file->f_dentry->d_inode;
512 struct ccc_object *vob = cl_inode2ccc(inode);
515 LASSERT(vma->vm_file);
516 LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
517 cfs_atomic_inc(&vob->cob_mmap_cnt);
522 * Dual to ll_vm_open().
524 static void ll_vm_close(struct vm_area_struct *vma)
526 struct inode *inode = vma->vm_file->f_dentry->d_inode;
527 struct ccc_object *vob = cl_inode2ccc(inode);
530 LASSERT(vma->vm_file);
531 cfs_atomic_dec(&vob->cob_mmap_cnt);
532 LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
536 #ifndef HAVE_VM_OP_FAULT
537 #ifndef HAVE_FILEMAP_POPULATE
538 static int (*filemap_populate)(struct vm_area_struct * area, unsigned long address, unsigned long len, pgprot_t prot, unsigned long pgoff, int nonblock);
540 static int ll_populate(struct vm_area_struct *area, unsigned long address,
541 unsigned long len, pgprot_t prot, unsigned long pgoff,
547 /* always set nonblock as true to avoid page read ahead */
548 rc = filemap_populate(area, address, len, prot, pgoff, 1);
553 /* return the user space pointer that maps to a file offset via a vma */
554 static inline unsigned long file_to_user(struct vm_area_struct *vma, __u64 byte)
556 return vma->vm_start + (byte - ((__u64)vma->vm_pgoff << CFS_PAGE_SHIFT));
560 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
561 * nopage's reference passing to the pte */
562 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
567 LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
568 if (mapping_mapped(mapping)) {
570 unmap_mapping_range(mapping, first + CFS_PAGE_SIZE - 1,
571 last - first + 1, 0);
577 static struct vm_operations_struct ll_file_vm_ops = {
578 #ifndef HAVE_VM_OP_FAULT
580 .populate = ll_populate,
585 .page_mkwrite = ll_page_mkwrite,
587 .close = ll_vm_close,
590 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
592 struct inode *inode = file->f_dentry->d_inode;
596 if (ll_file_nolock(file))
599 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
600 rc = generic_file_mmap(file, vma);
602 #if !defined(HAVE_FILEMAP_POPULATE) && !defined(HAVE_VM_OP_FAULT)
603 if (!filemap_populate)
604 filemap_populate = vma->vm_ops->populate;
606 vma->vm_ops = &ll_file_vm_ops;
607 vma->vm_ops->open(vma);
608 /* update the inode's size and mtime */
609 rc = ll_glimpse_size(inode);