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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #include <linux/errno.h>
34 #include <linux/delay.h>
35 #include <linux/kernel.h>
38 #define DEBUG_SUBSYSTEM S_LLITE
40 #include "llite_internal.h"
41 #include <lustre_compat.h>
43 static const struct vm_operations_struct ll_file_vm_ops;
45 void policy_from_vma(union ldlm_policy_data *policy, struct vm_area_struct *vma,
46 unsigned long addr, size_t count)
48 policy->l_extent.start = ((addr - vma->vm_start) & PAGE_MASK) +
49 (vma->vm_pgoff << PAGE_SHIFT);
50 policy->l_extent.end = (policy->l_extent.start + count - 1) |
54 struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
57 struct vm_area_struct *vma, *ret = NULL;
60 /* mmap_sem must have been held by caller. */
61 LASSERT(!down_write_trylock(&mm->mmap_sem));
63 for(vma = find_vma(mm, addr);
64 vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
65 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
66 vma->vm_flags & VM_SHARED) {
75 * API independent part for page fault initialization.
76 * \param env - corespondent lu_env to processing
77 * \param vma - virtual memory area addressed to page fault
78 * \param index - page index corespondent to fault.
79 * \parm ra_flags - vma readahead flags.
81 * \return error codes from cl_io_init.
84 ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma,
85 pgoff_t index, unsigned long *ra_flags)
87 struct file *file = vma->vm_file;
88 struct inode *inode = file_inode(file);
90 struct cl_fault_io *fio;
94 if (ll_file_nolock(file))
95 RETURN(ERR_PTR(-EOPNOTSUPP));
98 io = vvp_env_thread_io(env);
99 io->ci_obj = ll_i2info(inode)->lli_clob;
100 LASSERT(io->ci_obj != NULL);
102 fio = &io->u.ci_fault;
103 fio->ft_index = index;
104 fio->ft_executable = vma->vm_flags&VM_EXEC;
107 * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
108 * the kernel will not read other pages not covered by ldlm in
109 * filemap_nopage. we do our readahead in ll_readpage.
111 if (ra_flags != NULL)
112 *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
113 vma->vm_flags &= ~VM_SEQ_READ;
114 vma->vm_flags |= VM_RAND_READ;
116 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
117 fio->ft_index, fio->ft_executable);
119 rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
121 struct vvp_io *vio = vvp_env_io(env);
122 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
124 LASSERT(vio->vui_cl.cis_io == io);
126 /* mmap lock must be MANDATORY it has to cache
128 io->ci_lockreq = CILR_MANDATORY;
133 if (io->ci_need_restart)
142 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
143 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
152 struct inode *inode = NULL;
153 struct ll_inode_info *lli;
156 LASSERT(vmpage != NULL);
157 env = cl_env_get(&refcheck);
159 RETURN(PTR_ERR(env));
161 io = ll_fault_io_init(env, vma, vmpage->index, NULL);
163 GOTO(out, result = PTR_ERR(io));
165 result = io->ci_result;
167 GOTO(out_io, result);
169 io->u.ci_fault.ft_mkwrite = 1;
170 io->u.ci_fault.ft_writable = 1;
172 vio = vvp_env_io(env);
173 vio->u.fault.ft_vma = vma;
174 vio->u.fault.ft_vmpage = vmpage;
176 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
178 inode = vvp_object_inode(io->ci_obj);
179 lli = ll_i2info(inode);
181 result = cl_io_loop(env, io);
183 cfs_restore_sigs(set);
187 if (vmpage->mapping == NULL) {
190 /* page was truncated and lock was cancelled, return
191 * ENODATA so that VM_FAULT_NOPAGE will be returned
192 * to handle_mm_fault(). */
195 } else if (!PageDirty(vmpage)) {
196 /* race, the page has been cleaned by ptlrpcd after
197 * it was unlocked, it has to be added into dirty
198 * cache again otherwise this soon-to-dirty page won't
199 * consume any grants, even worse if this page is being
200 * transferred because it will break RPC checksum.
204 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
205 "been written out, retry.\n",
206 vmpage, vmpage->index);
213 ll_file_set_flag(lli, LLIF_DATA_MODIFIED);
220 cl_env_put(env, &refcheck);
221 CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
222 LASSERT(ergo(result == 0, PageLocked(vmpage)));
224 /* if page has been unmapped, presumably due to lock reclaim for
225 * concurrent usage, add some delay before retrying to prevent
226 * entering live-lock situation with competitors
228 if (result == -ENODATA && inode != NULL) {
229 CDEBUG(D_MMAP, "delaying new page-fault for inode %p to "
230 "prevent live-lock\n", inode);
237 static inline int to_fault_error(int result)
241 result = VM_FAULT_LOCKED;
244 result = VM_FAULT_OOM;
247 result = VM_FAULT_SIGBUS;
254 * Lustre implementation of a vm_operations_struct::fault() method, called by
255 * VM to server page fault (both in kernel and user space).
257 * \param vma - is virtiual area struct related to page fault
258 * \param vmf - structure which describe type and address where hit fault
260 * \return allocated and filled _locked_ page for address
261 * \retval VM_FAULT_ERROR on general error
262 * \retval NOPAGE_OOM not have memory for allocate new page
264 static vm_fault_t ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
268 struct vvp_io *vio = NULL;
270 unsigned long ra_flags;
276 env = cl_env_get(&refcheck);
278 RETURN(PTR_ERR(env));
280 if (ll_sbi_has_fast_read(ll_i2sbi(file_inode(vma->vm_file)))) {
282 ll_cl_add(vma->vm_file, env, NULL, LCC_MMAP);
283 fault_ret = ll_filemap_fault(vma, vmf);
284 ll_cl_remove(vma->vm_file, env);
286 /* - If there is no error, then the page was found in cache and
288 * - If VM_FAULT_RETRY is set, the page existed but failed to
289 * lock. It will return to kernel and retry;
290 * - Otherwise, it should try normal fault under DLM lock. */
291 if ((fault_ret & VM_FAULT_RETRY) ||
292 !(fault_ret & VM_FAULT_ERROR))
293 GOTO(out, result = 0);
298 io = ll_fault_io_init(env, vma, vmf->pgoff, &ra_flags);
300 GOTO(out, result = PTR_ERR(io));
302 result = io->ci_result;
304 vio = vvp_env_io(env);
305 vio->u.fault.ft_vma = vma;
306 vio->u.fault.ft_vmpage = NULL;
307 vio->u.fault.ft_vmf = vmf;
308 vio->u.fault.ft_flags = 0;
309 vio->u.fault.ft_flags_valid = 0;
311 /* May call ll_readpage() */
312 ll_cl_add(vma->vm_file, env, io, LCC_MMAP);
314 result = cl_io_loop(env, io);
316 ll_cl_remove(vma->vm_file, env);
318 /* ft_flags are only valid if we reached
319 * the call to filemap_fault */
320 if (vio->u.fault.ft_flags_valid)
321 fault_ret = vio->u.fault.ft_flags;
323 vmpage = vio->u.fault.ft_vmpage;
324 if (result != 0 && vmpage != NULL) {
331 vma->vm_flags |= ra_flags;
334 cl_env_put(env, &refcheck);
335 if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
336 fault_ret |= to_fault_error(result);
338 CDEBUG(D_MMAP, "%s fault %d/%d\n", current->comm, fault_ret, result);
342 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
343 static vm_fault_t ll_fault(struct vm_fault *vmf)
345 struct vm_area_struct *vma = vmf->vma;
347 static vm_fault_t ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
351 bool printed = false;
354 ktime_t kstart = ktime_get();
357 result = pcc_fault(vma, vmf, &cached);
361 /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
362 * so that it can be killed by admin but not cause segfault by
364 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
366 /* make sure offset is not a negative number */
367 if (vmf->pgoff > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
368 return VM_FAULT_SIGBUS;
370 result = ll_fault0(vma, vmf);
372 !(result & (VM_FAULT_RETRY | VM_FAULT_ERROR | VM_FAULT_LOCKED))) {
373 struct page *vmpage = vmf->page;
375 /* check if this page has been truncated */
377 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
382 if (!printed && ++count > 16) {
383 CWARN("the page is under heavy contention,"
384 "maybe your app(%s) needs revising :-)\n",
392 result |= VM_FAULT_LOCKED;
394 cfs_restore_sigs(set);
397 if (vmf->page && result == VM_FAULT_LOCKED) {
398 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
399 current->pid, LUSTRE_FPRIVATE(vma->vm_file),
400 cl_offset(NULL, vmf->page->index), PAGE_SIZE,
402 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
404 ktime_us_delta(ktime_get(), kstart));
410 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
411 static vm_fault_t ll_page_mkwrite(struct vm_fault *vmf)
413 struct vm_area_struct *vma = vmf->vma;
415 static vm_fault_t ll_page_mkwrite(struct vm_area_struct *vma,
416 struct vm_fault *vmf)
420 bool printed = false;
423 ktime_t kstart = ktime_get();
426 result = pcc_page_mkwrite(vma, vmf, &cached);
430 file_update_time(vma->vm_file);
433 result = ll_page_mkwrite0(vma, vmf->page, &retry);
435 if (!printed && ++count > 16) {
436 const struct dentry *de = file_dentry(vma->vm_file);
438 CWARN("app(%s): the page %lu of file "DFID" is under"
439 " heavy contention\n",
440 current->comm, vmf->pgoff,
441 PFID(ll_inode2fid(de->d_inode)));
448 LASSERT(PageLocked(vmf->page));
449 result = VM_FAULT_LOCKED;
453 result = VM_FAULT_NOPAGE;
456 result = VM_FAULT_OOM;
459 result = VM_FAULT_RETRY;
462 result = VM_FAULT_SIGBUS;
467 if (result == VM_FAULT_LOCKED) {
468 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
469 current->pid, LUSTRE_FPRIVATE(vma->vm_file),
470 cl_offset(NULL, vmf->page->index), PAGE_SIZE,
472 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
474 ktime_us_delta(ktime_get(), kstart));
481 * To avoid cancel the locks covering mmapped region for lock cache pressure,
482 * we track the mapped vma count in vvp_object::vob_mmap_cnt.
484 static void ll_vm_open(struct vm_area_struct * vma)
486 struct inode *inode = file_inode(vma->vm_file);
487 struct vvp_object *vob = cl_inode2vvp(inode);
490 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
491 atomic_inc(&vob->vob_mmap_cnt);
497 * Dual to ll_vm_open().
499 static void ll_vm_close(struct vm_area_struct *vma)
501 struct inode *inode = file_inode(vma->vm_file);
502 struct vvp_object *vob = cl_inode2vvp(inode);
505 atomic_dec(&vob->vob_mmap_cnt);
506 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
511 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
512 * nopage's reference passing to the pte */
513 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
518 LASSERTF(last > first, "last %llu first %llu\n", last, first);
519 if (mapping_mapped(mapping)) {
521 unmap_mapping_range(mapping, first + PAGE_SIZE - 1,
522 last - first + 1, 1);
528 static const struct vm_operations_struct ll_file_vm_ops = {
530 .page_mkwrite = ll_page_mkwrite,
532 .close = ll_vm_close,
535 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
537 struct inode *inode = file_inode(file);
538 ktime_t kstart = ktime_get();
544 if (ll_file_nolock(file))
547 rc = pcc_file_mmap(file, vma, &cached);
548 if (cached && rc != 0)
551 rc = generic_file_mmap(file, vma);
553 vma->vm_ops = &ll_file_vm_ops;
554 vma->vm_ops->open(vma);
555 /* update the inode's size and mtime */
557 rc = ll_glimpse_size(inode);
561 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MMAP,
562 ktime_us_delta(ktime_get(), kstart));