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, 2015, Intel Corporation.
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/uaccess.h>
46 #include <linux/pagemap.h>
48 #define DEBUG_SUBSYSTEM S_LLITE
50 #include "../include/lustre_lite.h"
51 #include "llite_internal.h"
52 #include "../include/linux/lustre_compat25.h"
54 static const struct vm_operations_struct ll_file_vm_ops;
56 void policy_from_vma(ldlm_policy_data_t *policy,
57 struct vm_area_struct *vma, unsigned long addr,
60 policy->l_extent.start = ((addr - vma->vm_start) & PAGE_MASK) +
61 (vma->vm_pgoff << PAGE_SHIFT);
62 policy->l_extent.end = (policy->l_extent.start + count - 1) |
66 struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
69 struct vm_area_struct *vma, *ret = NULL;
71 /* mmap_sem must have been held by caller. */
72 LASSERT(!down_write_trylock(&mm->mmap_sem));
74 for (vma = find_vma(mm, addr);
75 vma && vma->vm_start < (addr + count); vma = vma->vm_next) {
76 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
77 vma->vm_flags & VM_SHARED) {
86 * API independent part for page fault initialization.
87 * \param vma - virtual memory area addressed to page fault
88 * \param env - corespondent lu_env to processing
89 * \param nest - nested level
90 * \param index - page index corespondent to fault.
91 * \parm ra_flags - vma readahead flags.
93 * \return allocated and initialized env for fault operation.
94 * \retval EINVAL if env can't allocated
95 * \return other error codes from cl_io_init.
98 ll_fault_io_init(struct vm_area_struct *vma, struct lu_env **env_ret,
99 struct cl_env_nest *nest, pgoff_t index,
100 unsigned long *ra_flags)
102 struct file *file = vma->vm_file;
103 struct inode *inode = file_inode(file);
105 struct cl_fault_io *fio;
110 if (ll_file_nolock(file))
111 return ERR_PTR(-EOPNOTSUPP);
114 * page fault can be called when lustre IO is
115 * already active for the current thread, e.g., when doing read/write
116 * against user level buffer mapped from Lustre buffer. To avoid
117 * stomping on existing context, optionally force an allocation of a new
120 env = cl_env_nested_get(nest);
122 return ERR_PTR(-EINVAL);
127 io = vvp_env_thread_io(env);
128 io->ci_obj = ll_i2info(inode)->lli_clob;
131 fio = &io->u.ci_fault;
132 fio->ft_index = index;
133 fio->ft_executable = vma->vm_flags&VM_EXEC;
136 * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
137 * the kernel will not read other pages not covered by ldlm in
138 * filemap_nopage. we do our readahead in ll_readpage.
141 *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
142 vma->vm_flags &= ~VM_SEQ_READ;
143 vma->vm_flags |= VM_RAND_READ;
145 CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
146 fio->ft_index, fio->ft_executable);
148 rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
150 struct vvp_io *vio = vvp_env_io(env);
151 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
153 LASSERT(vio->vui_cl.cis_io == io);
155 /* mmap lock must be MANDATORY it has to cache pages. */
156 io->ci_lockreq = CILR_MANDATORY;
161 if (io->ci_need_restart)
164 cl_env_nested_put(nest, env);
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;
184 io = ll_fault_io_init(vma, &env, &nest, vmpage->index, NULL);
186 result = PTR_ERR(io);
190 result = io->ci_result;
194 io->u.ci_fault.ft_mkwrite = 1;
195 io->u.ci_fault.ft_writable = 1;
197 vio = vvp_env_io(env);
198 vio->u.fault.ft_vma = vma;
199 vio->u.fault.ft_vmpage = vmpage;
201 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
203 /* we grab lli_trunc_sem to exclude truncate case.
204 * Otherwise, we could add dirty pages into osc cache
205 * while truncate is on-going.
207 inode = vvp_object_inode(io->ci_obj);
208 lli = ll_i2info(inode);
209 down_read(&lli->lli_trunc_sem);
211 result = cl_io_loop(env, io);
213 up_read(&lli->lli_trunc_sem);
215 cfs_restore_sigs(set);
218 struct inode *inode = file_inode(vma->vm_file);
219 struct ll_inode_info *lli = ll_i2info(inode);
222 if (!vmpage->mapping) {
225 /* page was truncated and lock was cancelled, return
226 * ENODATA so that VM_FAULT_NOPAGE will be returned
227 * to handle_mm_fault().
231 } else if (!PageDirty(vmpage)) {
232 /* race, the page has been cleaned by ptlrpcd after
233 * it was unlocked, it has to be added into dirty
234 * cache again otherwise this soon-to-dirty page won't
235 * consume any grants, even worse if this page is being
236 * transferred because it will break RPC checksum.
240 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has been written out, retry.\n",
241 vmpage, vmpage->index);
248 spin_lock(&lli->lli_lock);
249 lli->lli_flags |= LLIF_DATA_MODIFIED;
250 spin_unlock(&lli->lli_lock);
256 cl_env_nested_put(&nest, env);
258 CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
259 LASSERT(ergo(result == 0, PageLocked(vmpage)));
264 static inline int to_fault_error(int result)
268 result = VM_FAULT_LOCKED;
271 result = VM_FAULT_NOPAGE;
274 result = VM_FAULT_OOM;
277 result = VM_FAULT_SIGBUS;
284 * Lustre implementation of a vm_operations_struct::fault() method, called by
285 * VM to server page fault (both in kernel and user space).
287 * \param vma - is virtual area struct related to page fault
288 * \param vmf - structure which describe type and address where hit fault
290 * \return allocated and filled _locked_ page for address
291 * \retval VM_FAULT_ERROR on general error
292 * \retval NOPAGE_OOM not have memory for allocate new page
294 static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
298 struct vvp_io *vio = NULL;
300 unsigned long ra_flags;
301 struct cl_env_nest nest;
305 io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags);
307 return to_fault_error(PTR_ERR(io));
309 result = io->ci_result;
311 vio = vvp_env_io(env);
312 vio->u.fault.ft_vma = vma;
313 vio->u.fault.ft_vmpage = NULL;
314 vio->u.fault.ft_vmf = vmf;
315 vio->u.fault.ft_flags = 0;
316 vio->u.fault.ft_flags_valid = false;
318 result = cl_io_loop(env, io);
320 /* ft_flags are only valid if we reached
321 * the call to filemap_fault
323 if (vio->u.fault.ft_flags_valid)
324 fault_ret = vio->u.fault.ft_flags;
326 vmpage = vio->u.fault.ft_vmpage;
327 if (result != 0 && vmpage) {
333 cl_env_nested_put(&nest, env);
335 vma->vm_flags |= ra_flags;
336 if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
337 fault_ret |= to_fault_error(result);
339 CDEBUG(D_MMAP, "%s fault %d/%d\n",
340 current->comm, fault_ret, result);
344 static int ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
347 bool printed = false;
351 /* Only SIGKILL and SIGTERM are allowed for fault/nopage/mkwrite
352 * so that it can be killed by admin but not cause segfault by
355 set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
358 result = ll_fault0(vma, vmf);
359 LASSERT(!(result & VM_FAULT_LOCKED));
361 struct page *vmpage = vmf->page;
363 /* check if this page has been truncated */
365 if (unlikely(!vmpage->mapping)) { /* unlucky */
370 if (!printed && ++count > 16) {
371 CWARN("the page is under heavy contention, maybe your app(%s) needs revising :-)\n",
379 result = VM_FAULT_LOCKED;
381 cfs_restore_sigs(set);
385 static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
388 bool printed = false;
394 result = ll_page_mkwrite0(vma, vmf->page, &retry);
396 if (!printed && ++count > 16) {
397 const struct dentry *de = vma->vm_file->f_path.dentry;
399 CWARN("app(%s): the page %lu of file "DFID" is under heavy contention\n",
400 current->comm, vmf->pgoff,
401 PFID(ll_inode2fid(de->d_inode)));
408 LASSERT(PageLocked(vmf->page));
409 result = VM_FAULT_LOCKED;
413 result = VM_FAULT_NOPAGE;
416 result = VM_FAULT_OOM;
419 result = VM_FAULT_RETRY;
422 result = VM_FAULT_SIGBUS;
430 * To avoid cancel the locks covering mmapped region for lock cache pressure,
431 * we track the mapped vma count in vvp_object::vob_mmap_cnt.
433 static void ll_vm_open(struct vm_area_struct *vma)
435 struct inode *inode = file_inode(vma->vm_file);
436 struct vvp_object *vob = cl_inode2vvp(inode);
438 LASSERT(vma->vm_file);
439 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
440 atomic_inc(&vob->vob_mmap_cnt);
444 * Dual to ll_vm_open().
446 static void ll_vm_close(struct vm_area_struct *vma)
448 struct inode *inode = file_inode(vma->vm_file);
449 struct vvp_object *vob = cl_inode2vvp(inode);
451 LASSERT(vma->vm_file);
452 atomic_dec(&vob->vob_mmap_cnt);
453 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
456 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
457 * nopage's reference passing to the pte
459 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
463 LASSERTF(last > first, "last %llu first %llu\n", last, first);
464 if (mapping_mapped(mapping)) {
466 unmap_mapping_range(mapping, first + PAGE_SIZE - 1,
467 last - first + 1, 0);
473 static const struct vm_operations_struct ll_file_vm_ops = {
475 .page_mkwrite = ll_page_mkwrite,
477 .close = ll_vm_close,
480 int ll_file_mmap(struct file *file, struct vm_area_struct *vma)
482 struct inode *inode = file_inode(file);
485 if (ll_file_nolock(file))
488 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
489 rc = generic_file_mmap(file, vma);
491 vma->vm_ops = &ll_file_vm_ops;
492 vma->vm_ops->open(vma);
493 /* update the inode's size and mtime */
494 rc = ll_glimpse_size(inode);