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/
32 #include <linux/errno.h>
33 #include <linux/delay.h>
34 #include <linux/kernel.h>
36 #include <linux/file.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) |
55 * Linux commit v6.0-rc3-225-gf39af05949a4
56 * mm: add VMA iterator
59 #define vma_iterator vm_area_struct *
60 #define vma_iter_init(vmip, mm, addr) *(vmip) = find_vma(mm, addr)
61 #define for_each_vma(vmi, vma) \
62 for (vma = vmi; vma != NULL; vma = vma->vm_next)
65 struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
68 struct vm_area_struct *vma, *ret = NULL;
69 struct vma_iterator vmi;
73 /* mmap_lock must have been held by caller. */
74 LASSERT(!mmap_write_trylock(mm));
76 vma_iter_init(&vmi, mm, addr);
77 for_each_vma(vmi, vma) {
78 if (vma->vm_start < (addr + count))
80 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
81 vma->vm_flags & VM_SHARED) {
90 * API independent part for page fault initialization.
91 * \param env - corespondent lu_env to processing
92 * \param vma - virtual memory area addressed to page fault
93 * \param index - page index corespondent to fault.
94 * \param mkwrite - whether it is mmap write.
96 * \return error codes from cl_io_init.
99 ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma,
100 pgoff_t index, bool mkwrite)
102 struct file *file = vma->vm_file;
103 struct inode *inode = file_inode(file);
105 struct cl_fault_io *fio;
109 if (ll_file_nolock(file))
110 RETURN(ERR_PTR(-EOPNOTSUPP));
113 io = vvp_env_thread_io(env);
114 io->ci_obj = ll_i2info(inode)->lli_clob;
115 LASSERT(io->ci_obj != NULL);
117 fio = &io->u.ci_fault;
118 fio->ft_index = index;
119 fio->ft_executable = vma->vm_flags & VM_EXEC;
123 fio->ft_writable = 1;
127 DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
128 PFID(&ll_i2info(inode)->lli_fid), vma, vma->vm_start,
129 vma->vm_end, vma->vm_flags, fio->ft_index);
131 if (vma->vm_flags & VM_SEQ_READ)
133 else if (vma->vm_flags & VM_RAND_READ)
134 io->ci_rand_read = 1;
136 rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
138 struct vvp_io *vio = vvp_env_io(env);
139 struct ll_file_data *fd = file->private_data;
141 LASSERT(vio->vui_cl.cis_io == io);
143 /* mmap lock must be MANDATORY it has to cache
145 io->ci_lockreq = CILR_MANDATORY;
149 if (io->ci_need_restart)
158 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
159 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
168 struct inode *inode = NULL;
169 struct ll_inode_info *lli;
172 LASSERT(vmpage != NULL);
173 env = cl_env_get(&refcheck);
175 RETURN(PTR_ERR(env));
177 io = ll_fault_io_init(env, vma, vmpage->index, true);
179 GOTO(out, result = PTR_ERR(io));
181 result = io->ci_result;
183 GOTO(out_io, result);
185 vio = vvp_env_io(env);
186 vio->u.fault.ft_vma = vma;
187 vio->u.fault.ft_vmpage = vmpage;
189 siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
190 sigprocmask(SIG_BLOCK, &new, &old);
192 inode = vvp_object_inode(io->ci_obj);
193 lli = ll_i2info(inode);
195 result = cl_io_loop(env, io);
197 sigprocmask(SIG_SETMASK, &old, NULL);
201 if (vmpage->mapping == NULL) {
204 /* page was truncated and lock was cancelled, return
205 * ENODATA so that VM_FAULT_NOPAGE will be returned
206 * to handle_mm_fault(). */
209 } else if (!PageDirty(vmpage)) {
210 /* race, the page has been cleaned by ptlrpcd after
211 * it was unlocked, it has to be added into dirty
212 * cache again otherwise this soon-to-dirty page won't
213 * consume any grants, even worse if this page is being
214 * transferred because it will break RPC checksum.
218 CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
219 "been written out, retry.\n",
220 vmpage, vmpage->index);
227 set_bit(LLIF_DATA_MODIFIED, &lli->lli_flags);
234 cl_env_put(env, &refcheck);
235 CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
236 LASSERT(ergo(result == 0, PageLocked(vmpage)));
238 /* if page has been unmapped, presumably due to lock reclaim for
239 * concurrent usage, add some delay before retrying to prevent
240 * entering live-lock situation with competitors
242 if (result == -ENODATA && inode != NULL) {
243 CDEBUG(D_MMAP, "delaying new page-fault for inode %p to "
244 "prevent live-lock\n", inode);
251 static inline int to_fault_error(int result)
255 result = VM_FAULT_LOCKED;
258 result = VM_FAULT_OOM;
261 result = VM_FAULT_SIGBUS;
267 int ll_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
269 struct inode *inode = file_inode(vma->vm_file);
273 /* this seqlock lets us notice if a page has been deleted on this inode
274 * during the fault process, allowing us to catch an erroneous SIGBUS
278 seq = read_seqbegin(&ll_i2info(inode)->lli_page_inv_lock);
279 ret = __ll_filemap_fault(vma, vmf);
280 } while (read_seqretry(&ll_i2info(inode)->lli_page_inv_lock, seq) &&
281 (ret & VM_FAULT_SIGBUS));
287 * Lustre implementation of a vm_operations_struct::fault() method, called by
288 * VM to server page fault (both in kernel and user space).
290 * \param vma - is virtiual area struct related to page fault
291 * \param vmf - structure which describe type and address where hit fault
293 * \return allocated and filled _locked_ page for address
294 * \retval VM_FAULT_ERROR on general error
295 * \retval NOPAGE_OOM not have memory for allocate new page
297 static vm_fault_t ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
299 struct inode *inode = file_inode(vma->vm_file);
302 struct vvp_io *vio = NULL;
309 env = cl_env_get(&refcheck);
311 RETURN(PTR_ERR(env));
313 if (ll_sbi_has_fast_read(ll_i2sbi(inode))) {
315 bool allow_retry = vmf->flags & FAULT_FLAG_ALLOW_RETRY;
316 bool has_retry = vmf->flags & FAULT_FLAG_RETRY_NOWAIT;
318 /* To avoid loops, instruct downstream to not drop mmap_sem */
320 * only need FAULT_FLAG_ALLOW_RETRY prior to Linux 5.1
321 * (6b4c9f4469819), where FAULT_FLAG_RETRY_NOWAIT is enough
322 * to not drop mmap_sem when failed to lock the page.
324 vmf->flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
325 ll_cl_add(inode, env, NULL, LCC_MMAP);
326 fault_ret = ll_filemap_fault(vma, vmf);
327 ll_cl_remove(inode, env);
329 vmf->flags &= ~FAULT_FLAG_RETRY_NOWAIT;
331 vmf->flags &= ~FAULT_FLAG_ALLOW_RETRY;
333 /* - If there is no error, then the page was found in cache and
335 * - If VM_FAULT_RETRY is set, the page existed but failed to
336 * lock. We will try slow path to avoid loops.
337 * - Otherwise, it should try normal fault under DLM lock. */
338 if (!(fault_ret & VM_FAULT_RETRY) &&
339 !(fault_ret & VM_FAULT_ERROR))
340 GOTO(out, result = 0);
345 io = ll_fault_io_init(env, vma, vmf->pgoff, false);
347 GOTO(out, result = PTR_ERR(io));
349 result = io->ci_result;
351 vio = vvp_env_io(env);
352 vio->u.fault.ft_vma = vma;
353 vio->u.fault.ft_vmpage = NULL;
354 vio->u.fault.ft_vmf = vmf;
355 vio->u.fault.ft_flags = 0;
356 vio->u.fault.ft_flags_valid = 0;
358 /* May call ll_readpage() */
359 ll_cl_add(inode, env, io, LCC_MMAP);
361 result = cl_io_loop(env, io);
363 ll_cl_remove(inode, env);
365 /* ft_flags are only valid if we reached
366 * the call to filemap_fault */
367 if (vio->u.fault.ft_flags_valid)
368 fault_ret = vio->u.fault.ft_flags;
370 vmpage = vio->u.fault.ft_vmpage;
371 if (result != 0 && vmpage != NULL) {
379 cl_env_put(env, &refcheck);
380 if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
381 fault_ret |= to_fault_error(result);
383 CDEBUG(D_MMAP, "%s fault %d/%d\n", current->comm, fault_ret, result);
387 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
388 static vm_fault_t ll_fault(struct vm_fault *vmf)
390 struct vm_area_struct *vma = vmf->vma;
392 static vm_fault_t ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
396 bool printed = false;
399 ktime_t kstart = ktime_get();
402 result = pcc_fault(vma, vmf, &cached);
406 CDEBUG(D_MMAP|D_IOTRACE,
407 DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
408 PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
409 vma, vma->vm_start, vma->vm_end, vma->vm_flags, vmf->pgoff);
411 /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
412 * so that it can be killed by admin but not cause segfault by
415 siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
416 sigprocmask(SIG_BLOCK, &new, &old);
418 /* make sure offset is not a negative number */
419 if (vmf->pgoff > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
420 return VM_FAULT_SIGBUS;
423 result = ll_fault0(vma, vmf);
425 !(result & (VM_FAULT_RETRY | VM_FAULT_ERROR | VM_FAULT_LOCKED))) {
426 struct page *vmpage = vmf->page;
428 /* lock the page, then check if this page has been truncated
429 * or deleted from Lustre and retry if so
432 if (unlikely(vmpage->mapping == NULL) ||
433 vmpage->private == 0) { /* unlucky */
438 if (!printed && ++count > 16) {
439 struct inode *inode = file_inode(vma->vm_file);
441 CWARN("%s: FID "DFID" under heavy mmap contention by '%s', consider revising IO pattern\n",
442 ll_i2sbi(inode)->ll_fsname,
443 PFID(&ll_i2info(inode)->lli_fid),
451 result |= VM_FAULT_LOCKED;
453 sigprocmask(SIG_SETMASK, &old, NULL);
456 if (vmf->page && result == VM_FAULT_LOCKED) {
457 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
458 current->pid, vma->vm_file->private_data,
459 cl_offset(NULL, vmf->page->index), PAGE_SIZE,
461 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
463 ktime_us_delta(ktime_get(), kstart));
467 "COMPLETED: "DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
468 PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
469 vma, vma->vm_start, vma->vm_end, vma->vm_flags, vmf->pgoff);
474 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
475 static vm_fault_t ll_page_mkwrite(struct vm_fault *vmf)
477 struct vm_area_struct *vma = vmf->vma;
479 static vm_fault_t ll_page_mkwrite(struct vm_area_struct *vma,
480 struct vm_fault *vmf)
484 bool printed = false;
487 ktime_t kstart = ktime_get();
490 CDEBUG(D_MMAP|D_IOTRACE,
491 DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
492 PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
493 vma, vma->vm_start, vma->vm_end, vma->vm_flags,
496 result = pcc_page_mkwrite(vma, vmf, &cached);
500 file_update_time(vma->vm_file);
503 result = ll_page_mkwrite0(vma, vmf->page, &retry);
505 if (!printed && ++count > 16) {
506 const struct dentry *de = file_dentry(vma->vm_file);
508 CWARN("app(%s): the page %lu of file "DFID" is under heavy contention\n",
509 current->comm, vmf->pgoff,
510 PFID(ll_inode2fid(de->d_inode)));
517 LASSERT(PageLocked(vmf->page));
518 result = VM_FAULT_LOCKED;
522 result = VM_FAULT_NOPAGE;
525 result = VM_FAULT_OOM;
528 result = VM_FAULT_RETRY;
531 result = VM_FAULT_SIGBUS;
536 if (result == VM_FAULT_LOCKED) {
537 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
538 current->pid, vma->vm_file->private_data,
539 cl_offset(NULL, vmf->page->index), PAGE_SIZE,
541 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
543 ktime_us_delta(ktime_get(), kstart));
547 "COMPLETED: "DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
548 PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
549 vma, vma->vm_start, vma->vm_end, vma->vm_flags,
555 * To avoid cancel the locks covering mmapped region for lock cache pressure,
556 * we track the mapped vma count in vvp_object::vob_mmap_cnt.
558 static void ll_vm_open(struct vm_area_struct * vma)
560 struct inode *inode = file_inode(vma->vm_file);
561 struct vvp_object *vob = cl_inode2vvp(inode);
564 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
565 atomic_inc(&vob->vob_mmap_cnt);
571 * Dual to ll_vm_open().
573 static void ll_vm_close(struct vm_area_struct *vma)
575 struct inode *inode = file_inode(vma->vm_file);
576 struct vvp_object *vob = cl_inode2vvp(inode);
579 atomic_dec(&vob->vob_mmap_cnt);
580 LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
585 static const struct vm_operations_struct ll_file_vm_ops = {
587 .page_mkwrite = ll_page_mkwrite,
589 .close = ll_vm_close,
592 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
594 struct inode *inode = file_inode(file);
595 ktime_t kstart = ktime_get();
600 CDEBUG(D_VFSTRACE | D_MMAP,
601 "VFS_Op: fid="DFID" vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
602 PFID(&ll_i2info(inode)->lli_fid),
603 vma, vma->vm_start, vma->vm_end, vma->vm_flags);
605 if (ll_file_nolock(file))
608 rc = pcc_file_mmap(file, vma, &cached);
609 if (cached && rc != 0)
612 rc = generic_file_mmap(file, vma);
614 vma->vm_ops = &ll_file_vm_ops;
615 vma->vm_ops->open(vma);
616 /* update the inode's size and mtime */
618 rc = ll_glimpse_size(inode);
622 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MMAP,
623 ktime_us_delta(ktime_get(), kstart));