1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
40 #ifndef AUTOCONF_INCLUDED
41 #include <linux/config.h>
43 #include <linux/kernel.h>
45 #include <linux/string.h>
46 #include <linux/stat.h>
47 #include <linux/errno.h>
48 #include <linux/smp_lock.h>
49 #include <linux/unistd.h>
50 #include <linux/version.h>
51 #include <asm/system.h>
52 #include <asm/uaccess.h>
55 #include <linux/stat.h>
56 #include <asm/uaccess.h>
58 #include <linux/pagemap.h>
59 #include <linux/smp_lock.h>
61 #define DEBUG_SUBSYSTEM S_LLITE
63 #include <lustre_lite.h>
64 #include "llite_internal.h"
65 #include <linux/lustre_compat25.h>
67 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
70 static struct vm_operations_struct ll_file_vm_ops;
72 void policy_from_vma(ldlm_policy_data_t *policy,
73 struct vm_area_struct *vma, unsigned long addr,
76 policy->l_extent.start = ((addr - vma->vm_start) & CFS_PAGE_MASK) +
77 (vma->vm_pgoff << CFS_PAGE_SHIFT);
78 policy->l_extent.end = (policy->l_extent.start + count - 1) |
82 struct vm_area_struct * our_vma(unsigned long addr, size_t count)
84 struct mm_struct *mm = current->mm;
85 struct vm_area_struct *vma, *ret = NULL;
88 /* No MM (e.g. NFS)? No vmas too. */
92 spin_lock(&mm->page_table_lock);
93 for(vma = find_vma(mm, addr);
94 vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
95 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
96 vma->vm_flags & VM_SHARED) {
101 spin_unlock(&mm->page_table_lock);
106 * API independent part for page fault initialization.
107 * \param vma - virtual memory area addressed to page fault
108 * \param env - corespondent lu_env to processing
109 * \param nest - nested level
110 * \param index - page index corespondent to fault.
111 * \parm ra_flags - vma readahead flags.
113 * \return allocated and initialized env for fault operation.
114 * \retval EINVAL if env can't allocated
115 * \return other error codes from cl_io_init.
117 struct cl_io *ll_fault_io_init(struct vm_area_struct *vma,
118 struct lu_env **env_ret,
119 struct cl_env_nest *nest,
120 pgoff_t index, unsigned long *ra_flags)
122 struct file *file = vma->vm_file;
123 struct inode *inode = file->f_dentry->d_inode;
125 struct cl_fault_io *fio;
130 if (ll_file_nolock(file))
131 RETURN(ERR_PTR(-EOPNOTSUPP));
134 * page fault can be called when lustre IO is
135 * already active for the current thread, e.g., when doing read/write
136 * against user level buffer mapped from Lustre buffer. To avoid
137 * stomping on existing context, optionally force an allocation of a new
140 env = cl_env_nested_get(nest);
142 RETURN(ERR_PTR(-EINVAL));
146 io = ccc_env_thread_io(env);
147 io->ci_obj = ll_i2info(inode)->lli_clob;
148 LASSERT(io->ci_obj != NULL);
150 fio = &io->u.ci_fault;
151 fio->ft_index = index;
152 fio->ft_executable = vma->vm_flags&VM_EXEC;
155 * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
156 * the kernel will not read other pages not covered by ldlm in
157 * filemap_nopage. we do our readahead in ll_readpage.
159 if (ra_flags != NULL)
160 *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
161 vma->vm_flags &= ~VM_SEQ_READ;
162 vma->vm_flags |= VM_RAND_READ;
164 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
165 fio->ft_index, fio->ft_executable);
167 if (cl_io_init(env, io, CIT_FAULT, io->ci_obj) == 0) {
168 struct ccc_io *cio = ccc_env_io(env);
169 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
171 LASSERT(cio->cui_cl.cis_io == io);
173 /* mmap lock must be MANDATORY
174 * it has to cache pages. */
175 io->ci_lockreq = CILR_MANDATORY;
183 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
184 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
190 struct cl_env_nest nest;
194 LASSERT(vmpage != NULL);
196 io = ll_fault_io_init(vma, &env, &nest, vmpage->index, NULL);
198 GOTO(out, result = PTR_ERR(io));
200 result = io->ci_result;
204 /* Don't enqueue new locks for page_mkwrite().
205 * If the lock has been cancelled then page must have been
206 * truncated, in that case, kernel will handle it.
208 io->ci_lockreq = CILR_PEEK;
209 io->u.ci_fault.ft_mkwrite = 1;
210 io->u.ci_fault.ft_writable = 1;
212 vio = vvp_env_io(env);
213 vio->u.fault.ft_vma = vma;
214 vio->u.fault.ft_vmpage = vmpage;
216 result = cl_io_loop(env, io);
218 if (result == -ENODATA) /* peek failed, no lock caching. */
219 CDEBUG(D_MMAP, "race on page_mkwrite: %lx (%lu %p)\n",
220 vma->vm_flags, io->u.ci_fault.ft_index, vmpage);
222 if (result == 0 || result == -ENODATA) {
224 if (vmpage->mapping == NULL) {
227 /* page was truncated and lock was cancelled, return
228 * ENODATA so that VM_FAULT_NOPAGE will be returned
229 * to handle_mm_fault(). */
232 } else if (result == -ENODATA) {
233 /* Invalidate it if the cl_lock is being revoked.
234 * This piece of code is definitely needed for RHEL5,
235 * otherwise, SIGBUS will be wrongly returned to
237 ll_invalidate_page(vmpage);
238 LASSERT(vmpage->mapping == NULL);
240 } else if (!PageDirty(vmpage)) {
241 /* race, the page has been cleaned by ptlrpcd after
242 * it was unlocked, it has to be added into dirty
243 * cache again otherwise this soon-to-dirty page won't
244 * consume any grants, even worse if this page is being
245 * transferred because it will break RPC checksum.
249 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
250 "been written out, retry.\n",
251 vmpage, vmpage->index);
261 cl_env_nested_put(&nest, env);
263 CDEBUG(D_MMAP, "%s mkwrite with %d\n", cfs_current()->comm, result);
265 LASSERT(ergo(result == 0, PageLocked(vmpage)));
270 #ifndef HAVE_VM_OP_FAULT
272 * Lustre implementation of a vm_operations_struct::nopage() method, called by
273 * VM to server page fault (both in kernel and user space).
275 * This function sets up CIT_FAULT cl_io that does the job.
277 * \param vma - is virtiual area struct related to page fault
278 * \param address - address when hit fault
279 * \param type - of fault
281 * \return allocated and filled _unlocked_ page for address
282 * \retval NOPAGE_SIGBUS if page not exist on this address
283 * \retval NOPAGE_OOM not have memory for allocate new page
285 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
289 struct cl_env_nest nest;
291 struct page *page = NOPAGE_SIGBUS;
292 struct vvp_io *vio = NULL;
293 unsigned long ra_flags;
296 const unsigned long writable = VM_SHARED|VM_WRITE;
299 pg_offset = ((address - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
300 io = ll_fault_io_init(vma, &env, &nest, pg_offset, &ra_flags);
302 return NOPAGE_SIGBUS;
304 result = io->ci_result;
308 io->u.ci_fault.ft_writable = (vma->vm_flags&writable) == writable;
310 vio = vvp_env_io(env);
311 vio->u.fault.ft_vma = vma;
312 vio->u.fault.nopage.ft_address = address;
313 vio->u.fault.nopage.ft_type = type;
314 vio->u.fault.ft_vmpage = NULL;
316 result = cl_io_loop(env, io);
317 page = vio->u.fault.ft_vmpage;
318 if (result != 0 && page != NULL)
319 page_cache_release(page);
322 if (result == -ENOMEM)
325 vma->vm_flags &= ~VM_RAND_READ;
326 vma->vm_flags |= ra_flags;
329 cl_env_nested_put(&nest, env);
334 static int ll_page_mkwrite(struct vm_area_struct *vma, struct page *vmpage)
337 bool printed = false;
343 result = ll_page_mkwrite0(vma, vmpage, &retry);
345 if (!printed && ++count > 16) {
346 CWARN("app(%s): the page %lu of file %lu is under heavy"
348 current->comm, page_index(vmpage),
349 vma->vm_file->f_dentry->d_inode->i_ino);
356 else if (result == -ENODATA)
357 result = 0; /* kernel will know truncate has happened and
365 * Lustre implementation of a vm_operations_struct::fault() method, called by
366 * VM to server page fault (both in kernel and user space).
368 * \param vma - is virtiual area struct related to page fault
369 * \param vmf - structure which describe type and address where hit fault
371 * \return allocated and filled _locked_ page for address
372 * \retval VM_FAULT_ERROR on general error
373 * \retval NOPAGE_OOM not have memory for allocate new page
375 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
379 struct vvp_io *vio = NULL;
381 unsigned long ra_flags;
382 struct cl_env_nest nest;
387 io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags);
389 RETURN(VM_FAULT_ERROR);
391 result = io->ci_result;
395 vio = vvp_env_io(env);
396 vio->u.fault.ft_vma = vma;
397 vio->u.fault.ft_vmpage = NULL;
398 vio->u.fault.fault.ft_vmf = vmf;
400 result = cl_io_loop(env, io);
402 vmpage = vio->u.fault.ft_vmpage;
403 if (result != 0 && vmpage != NULL) {
404 page_cache_release(vmpage);
408 fault_ret = vio->u.fault.fault.ft_flags;
411 if (result != 0 && fault_ret == 0)
412 fault_ret = VM_FAULT_ERROR;
414 vma->vm_flags |= ra_flags;
417 cl_env_nested_put(&nest, env);
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 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;
458 static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
461 bool printed = false;
467 result = ll_page_mkwrite0(vma, vmf->page, &retry);
469 if (!printed && ++count > 16) {
470 CWARN("app(%s): the page %lu of file %lu is under heavy"
472 current->comm, vmf->pgoff,
473 vma->vm_file->f_dentry->d_inode->i_ino);
480 LASSERT(PageLocked(vmf->page));
481 result = VM_FAULT_LOCKED;
485 result = VM_FAULT_NOPAGE;
488 result = VM_FAULT_OOM;
491 result = VM_FAULT_RETRY;
494 result = VM_FAULT_SIGBUS;
503 * To avoid cancel the locks covering mmapped region for lock cache pressure,
504 * we track the mapped vma count in ccc_object::cob_mmap_cnt.
506 static void ll_vm_open(struct vm_area_struct * vma)
508 struct inode *inode = vma->vm_file->f_dentry->d_inode;
509 struct ccc_object *vob = cl_inode2ccc(inode);
512 LASSERT(vma->vm_file);
513 LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
514 cfs_atomic_inc(&vob->cob_mmap_cnt);
519 * Dual to ll_vm_open().
521 static void ll_vm_close(struct vm_area_struct *vma)
523 struct inode *inode = vma->vm_file->f_dentry->d_inode;
524 struct ccc_object *vob = cl_inode2ccc(inode);
527 LASSERT(vma->vm_file);
528 cfs_atomic_dec(&vob->cob_mmap_cnt);
529 LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
533 #ifndef HAVE_VM_OP_FAULT
534 #ifndef HAVE_FILEMAP_POPULATE
535 static int (*filemap_populate)(struct vm_area_struct * area, unsigned long address, unsigned long len, pgprot_t prot, unsigned long pgoff, int nonblock);
537 static int ll_populate(struct vm_area_struct *area, unsigned long address,
538 unsigned long len, pgprot_t prot, unsigned long pgoff,
544 /* always set nonblock as true to avoid page read ahead */
545 rc = filemap_populate(area, address, len, prot, pgoff, 1);
550 /* return the user space pointer that maps to a file offset via a vma */
551 static inline unsigned long file_to_user(struct vm_area_struct *vma, __u64 byte)
553 return vma->vm_start + (byte - ((__u64)vma->vm_pgoff << CFS_PAGE_SHIFT));
557 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
558 * nopage's reference passing to the pte */
559 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
564 LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
565 if (mapping_mapped(mapping)) {
567 unmap_mapping_range(mapping, first + CFS_PAGE_SIZE - 1,
568 last - first + 1, 0);
574 static struct vm_operations_struct ll_file_vm_ops = {
575 #ifndef HAVE_VM_OP_FAULT
577 .populate = ll_populate,
582 .page_mkwrite = ll_page_mkwrite,
584 .close = ll_vm_close,
587 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
589 struct inode *inode = file->f_dentry->d_inode;
593 if (ll_file_nolock(file))
596 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
597 rc = generic_file_mmap(file, vma);
599 #if !defined(HAVE_FILEMAP_POPULATE) && !defined(HAVE_VM_OP_FAULT)
600 if (!filemap_populate)
601 filemap_populate = vma->vm_ops->populate;
603 vma->vm_ops = &ll_file_vm_ops;
604 vma->vm_ops->open(vma);
605 /* update the inode's size and mtime */
606 rc = cl_glimpse_size(inode);