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(struct mm_struct *mm, unsigned long addr,
82 struct vm_area_struct *vma, *ret = NULL;
85 /* mmap_sem must have been held by caller. */
86 LASSERT(!down_write_trylock(&mm->mmap_sem));
88 for(vma = find_vma(mm, addr);
89 vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
90 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
91 vma->vm_flags & VM_SHARED) {
100 * API independent part for page fault initialization.
101 * \param vma - virtual memory area addressed to page fault
102 * \param env - corespondent lu_env to processing
103 * \param nest - nested level
104 * \param index - page index corespondent to fault.
105 * \parm ra_flags - vma readahead flags.
107 * \return allocated and initialized env for fault operation.
108 * \retval EINVAL if env can't allocated
109 * \return other error codes from cl_io_init.
111 struct cl_io *ll_fault_io_init(struct vm_area_struct *vma,
112 struct lu_env **env_ret,
113 struct cl_env_nest *nest,
114 pgoff_t index, unsigned long *ra_flags)
116 struct file *file = vma->vm_file;
117 struct inode *inode = file->f_dentry->d_inode;
119 struct cl_fault_io *fio;
124 if (ll_file_nolock(file))
125 RETURN(ERR_PTR(-EOPNOTSUPP));
128 * page fault can be called when lustre IO is
129 * already active for the current thread, e.g., when doing read/write
130 * against user level buffer mapped from Lustre buffer. To avoid
131 * stomping on existing context, optionally force an allocation of a new
134 env = cl_env_nested_get(nest);
136 RETURN(ERR_PTR(-EINVAL));
140 io = ccc_env_thread_io(env);
141 io->ci_obj = ll_i2info(inode)->lli_clob;
142 LASSERT(io->ci_obj != NULL);
144 fio = &io->u.ci_fault;
145 fio->ft_index = index;
146 fio->ft_executable = vma->vm_flags&VM_EXEC;
149 * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
150 * the kernel will not read other pages not covered by ldlm in
151 * filemap_nopage. we do our readahead in ll_readpage.
153 if (ra_flags != NULL)
154 *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
155 vma->vm_flags &= ~VM_SEQ_READ;
156 vma->vm_flags |= VM_RAND_READ;
158 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
159 fio->ft_index, fio->ft_executable);
161 if (cl_io_init(env, io, CIT_FAULT, io->ci_obj) == 0) {
162 struct ccc_io *cio = ccc_env_io(env);
163 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
165 LASSERT(cio->cui_cl.cis_io == io);
167 /* mmap lock must be MANDATORY
168 * it has to cache pages. */
169 io->ci_lockreq = CILR_MANDATORY;
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;
189 LASSERT(vmpage != NULL);
191 io = ll_fault_io_init(vma, &env, &nest, vmpage->index, NULL);
193 GOTO(out, result = PTR_ERR(io));
195 result = io->ci_result;
199 /* Don't enqueue new locks for page_mkwrite().
200 * If the lock has been cancelled then page must have been
201 * truncated, in that case, kernel will handle it.
203 io->ci_lockreq = CILR_PEEK;
204 io->u.ci_fault.ft_mkwrite = 1;
205 io->u.ci_fault.ft_writable = 1;
207 vio = vvp_env_io(env);
208 vio->u.fault.ft_vma = vma;
209 vio->u.fault.ft_vmpage = vmpage;
211 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
212 result = cl_io_loop(env, io);
213 cfs_restore_sigs(set);
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;
301 pg_offset = ((address - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
302 io = ll_fault_io_init(vma, &env, &nest, pg_offset, &ra_flags);
304 return NOPAGE_SIGBUS;
306 result = io->ci_result;
310 io->u.ci_fault.ft_writable = (vma->vm_flags&writable) == writable;
312 vio = vvp_env_io(env);
313 vio->u.fault.ft_vma = vma;
314 vio->u.fault.nopage.ft_address = address;
315 vio->u.fault.nopage.ft_type = type;
316 vio->u.fault.ft_vmpage = NULL;
318 set = cfs_block_sigsinv(sigmask(SIGKILL)|sigmask(SIGTERM));
319 result = cl_io_loop(env, io);
320 cfs_restore_sigs(set);
322 page = vio->u.fault.ft_vmpage;
323 if (result != 0 && page != NULL) {
324 page_cache_release(page);
325 page = NOPAGE_SIGBUS;
329 if (result == -ENOMEM)
332 vma->vm_flags &= ~VM_RAND_READ;
333 vma->vm_flags |= ra_flags;
336 cl_env_nested_put(&nest, env);
343 static inline int to_fault_error(int result)
347 result = VM_FAULT_LOCKED;
350 result = VM_FAULT_NOPAGE;
353 result = VM_FAULT_OOM;
356 result = VM_FAULT_SIGBUS;
363 * Lustre implementation of a vm_operations_struct::fault() method, called by
364 * VM to server page fault (both in kernel and user space).
366 * \param vma - is virtiual area struct related to page fault
367 * \param vmf - structure which describe type and address where hit fault
369 * \return allocated and filled _locked_ page for address
370 * \retval VM_FAULT_ERROR on general error
371 * \retval NOPAGE_OOM not have memory for allocate new page
373 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
377 struct vvp_io *vio = NULL;
379 unsigned long ra_flags;
380 struct cl_env_nest nest;
385 io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags);
387 RETURN(to_fault_error(PTR_ERR(io)));
389 result = io->ci_result;
391 vio = vvp_env_io(env);
392 vio->u.fault.ft_vma = vma;
393 vio->u.fault.ft_vmpage = NULL;
394 vio->u.fault.fault.ft_vmf = vmf;
396 result = cl_io_loop(env, io);
398 fault_ret = vio->u.fault.fault.ft_flags;
399 vmpage = vio->u.fault.ft_vmpage;
400 if (result != 0 && vmpage != NULL) {
401 page_cache_release(vmpage);
406 cl_env_nested_put(&nest, env);
408 vma->vm_flags |= ra_flags;
409 if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
410 fault_ret |= to_fault_error(result);
412 CDEBUG(D_MMAP, "%s fault %d/%d\n",
413 cfs_current()->comm, fault_ret, result);
417 static int ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
420 bool printed = false;
424 /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
425 * so that it can be killed by admin but not cause segfault by
427 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
430 result = ll_fault0(vma, vmf);
431 LASSERT(!(result & VM_FAULT_LOCKED));
433 struct page *vmpage = vmf->page;
435 /* check if this page has been truncated */
437 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
439 page_cache_release(vmpage);
442 if (!printed && ++count > 16) {
443 CWARN("the page is under heavy contention,"
444 "maybe your app(%s) needs revising :-)\n",
452 result |= VM_FAULT_LOCKED;
454 cfs_restore_sigs(set);
459 #ifndef HAVE_PGMKWRITE_USE_VMFAULT
460 static int ll_page_mkwrite(struct vm_area_struct *vma, struct page *vmpage)
463 bool printed = false;
469 result = ll_page_mkwrite0(vma, vmpage, &retry);
471 if (!printed && ++count > 16) {
472 CWARN("app(%s): the page %lu of file %lu is under heavy"
474 current->comm, page_index(vmpage),
475 vma->vm_file->f_dentry->d_inode->i_ino);
482 else if (result == -ENODATA)
483 result = 0; /* kernel will know truncate has happened and
489 static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
492 bool printed = false;
498 result = ll_page_mkwrite0(vma, vmf->page, &retry);
500 if (!printed && ++count > 16) {
501 CWARN("app(%s): the page %lu of file %lu is under heavy"
503 current->comm, vmf->pgoff,
504 vma->vm_file->f_dentry->d_inode->i_ino);
511 LASSERT(PageLocked(vmf->page));
512 result = VM_FAULT_LOCKED;
516 result = VM_FAULT_NOPAGE;
519 result = VM_FAULT_OOM;
522 result = VM_FAULT_RETRY;
525 result = VM_FAULT_SIGBUS;
534 * To avoid cancel the locks covering mmapped region for lock cache pressure,
535 * we track the mapped vma count in ccc_object::cob_mmap_cnt.
537 static void ll_vm_open(struct vm_area_struct * vma)
539 struct inode *inode = vma->vm_file->f_dentry->d_inode;
540 struct ccc_object *vob = cl_inode2ccc(inode);
543 LASSERT(vma->vm_file);
544 LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
545 cfs_atomic_inc(&vob->cob_mmap_cnt);
550 * Dual to ll_vm_open().
552 static void ll_vm_close(struct vm_area_struct *vma)
554 struct inode *inode = vma->vm_file->f_dentry->d_inode;
555 struct ccc_object *vob = cl_inode2ccc(inode);
558 LASSERT(vma->vm_file);
559 cfs_atomic_dec(&vob->cob_mmap_cnt);
560 LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
564 #ifndef HAVE_VM_OP_FAULT
565 #ifndef HAVE_FILEMAP_POPULATE
566 static int (*filemap_populate)(struct vm_area_struct * area, unsigned long address, unsigned long len, pgprot_t prot, unsigned long pgoff, int nonblock);
568 static int ll_populate(struct vm_area_struct *area, unsigned long address,
569 unsigned long len, pgprot_t prot, unsigned long pgoff,
575 /* always set nonblock as true to avoid page read ahead */
576 rc = filemap_populate(area, address, len, prot, pgoff, 1);
581 /* return the user space pointer that maps to a file offset via a vma */
582 static inline unsigned long file_to_user(struct vm_area_struct *vma, __u64 byte)
584 return vma->vm_start + (byte - ((__u64)vma->vm_pgoff << CFS_PAGE_SHIFT));
588 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
589 * nopage's reference passing to the pte */
590 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
595 LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
596 if (mapping_mapped(mapping)) {
598 unmap_mapping_range(mapping, first + CFS_PAGE_SIZE - 1,
599 last - first + 1, 0);
605 static struct vm_operations_struct ll_file_vm_ops = {
606 #ifndef HAVE_VM_OP_FAULT
608 .populate = ll_populate,
613 .page_mkwrite = ll_page_mkwrite,
615 .close = ll_vm_close,
618 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
620 struct inode *inode = file->f_dentry->d_inode;
624 if (ll_file_nolock(file))
627 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
628 rc = generic_file_mmap(file, vma);
630 #if !defined(HAVE_FILEMAP_POPULATE) && !defined(HAVE_VM_OP_FAULT)
631 if (!filemap_populate)
632 filemap_populate = vma->vm_ops->populate;
634 vma->vm_ops = &ll_file_vm_ops;
635 vma->vm_ops->open(vma);
636 /* update the inode's size and mtime */
637 rc = ll_glimpse_size(inode);