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) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/lustre/llite/rw26.c
38 * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
41 #include <linux/kernel.h>
43 #include <linux/string.h>
44 #include <linux/stat.h>
45 #include <linux/errno.h>
46 #include <linux/unistd.h>
47 #include <asm/uaccess.h>
50 #include <linux/migrate.h>
51 #elif defined(HAVE_MIGRATE_MODE_H)
52 #include <linux/migrate_mode.h>
55 #include <linux/buffer_head.h>
56 #include <linux/mpage.h>
57 #include <linux/writeback.h>
58 #include <linux/stat.h>
59 #include <asm/uaccess.h>
61 #include <linux/pagemap.h>
63 #define DEBUG_SUBSYSTEM S_LLITE
65 #include "llite_internal.h"
66 #include <lustre_compat.h>
69 * Implements Linux VM address_space::invalidatepage() method. This method is
70 * called when the page is truncate from a file, either as a result of
71 * explicit truncate, or when inode is removed from memory (as a result of
72 * final iput(), umount, or memory pressure induced icache shrinking).
74 * [0, offset] bytes of the page remain valid (this is for a case of not-page
75 * aligned truncate). Lustre leaves partially truncated page in the cache,
76 * relying on struct inode::i_size to limit further accesses.
78 static void ll_invalidatepage(struct page *vmpage,
79 #ifdef HAVE_INVALIDATE_RANGE
80 unsigned int offset, unsigned int length
89 struct cl_object *obj;
93 LASSERT(PageLocked(vmpage));
94 LASSERT(!PageWriteback(vmpage));
97 * It is safe to not check anything in invalidatepage/releasepage
98 * below because they are run with page locked and all our io is
99 * happening with locked page too
101 #ifdef HAVE_INVALIDATE_RANGE
102 if (offset == 0 && length == PAGE_CACHE_SIZE) {
106 env = cl_env_get(&refcheck);
108 inode = vmpage->mapping->host;
109 obj = ll_i2info(inode)->lli_clob;
111 page = cl_vmpage_page(vmpage, obj);
113 cl_page_delete(env, page);
114 cl_page_put(env, page);
117 LASSERT(vmpage->private == 0);
118 cl_env_put(env, &refcheck);
123 #ifdef HAVE_RELEASEPAGE_WITH_INT
124 #define RELEASEPAGE_ARG_TYPE int
126 #define RELEASEPAGE_ARG_TYPE gfp_t
128 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
132 struct cl_object *obj;
133 struct cl_page *page;
134 struct address_space *mapping;
137 LASSERT(PageLocked(vmpage));
138 if (PageWriteback(vmpage) || PageDirty(vmpage))
141 mapping = vmpage->mapping;
145 obj = ll_i2info(mapping->host)->lli_clob;
149 /* 1 for caller, 1 for cl_page and 1 for page cache */
150 if (page_count(vmpage) > 3)
153 page = cl_vmpage_page(vmpage, obj);
157 cookie = cl_env_reenter();
158 env = cl_env_percpu_get();
159 LASSERT(!IS_ERR(env));
161 if (!cl_page_in_use(page)) {
163 cl_page_delete(env, page);
166 /* To use percpu env array, the call path can not be rescheduled;
167 * otherwise percpu array will be messed if ll_releaspage() called
168 * again on the same CPU.
170 * If this page holds the last refc of cl_object, the following
171 * call path may cause reschedule:
172 * cl_page_put -> cl_page_free -> cl_object_put ->
173 * lu_object_put -> lu_object_free -> lov_delete_raid0.
175 * However, the kernel can't get rid of this inode until all pages have
176 * been cleaned up. Now that we hold page lock here, it's pretty safe
177 * that we won't get into object delete path.
179 LASSERT(cl_object_refc(obj) > 1);
180 cl_page_put(env, page);
182 cl_env_percpu_put(env);
183 cl_env_reexit(cookie);
187 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
189 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
190 size_t size, struct page ***pages,
193 int result = -ENOMEM;
195 /* set an arbitrary limit to prevent arithmetic overflow */
196 if (size > MAX_DIRECTIO_SIZE) {
201 *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >>
203 *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
205 OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
207 down_read(¤t->mm->mmap_sem);
208 result = get_user_pages(current, current->mm, user_addr,
209 *max_pages, (rw == READ), 0, *pages,
211 up_read(¤t->mm->mmap_sem);
212 if (unlikely(result <= 0))
213 OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
219 /* ll_free_user_pages - tear down page struct array
220 * @pages: array of page struct pointers underlying target buffer */
221 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
225 for (i = 0; i < npages; i++) {
226 if (pages[i] == NULL)
229 set_page_dirty_lock(pages[i]);
230 page_cache_release(pages[i]);
233 OBD_FREE_LARGE(pages, npages * sizeof(*pages));
236 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
237 int rw, struct inode *inode,
238 struct ll_dio_pages *pv)
241 struct cl_2queue *queue;
242 struct cl_object *obj = io->ci_obj;
245 loff_t file_offset = pv->ldp_start_offset;
246 size_t size = pv->ldp_size;
247 int page_count = pv->ldp_nr;
248 struct page **pages = pv->ldp_pages;
249 size_t page_size = cl_page_size(obj);
254 queue = &io->ci_queue;
255 cl_2queue_init(queue);
256 for (i = 0; i < page_count; i++) {
258 file_offset = pv->ldp_offsets[i];
260 LASSERT(!(file_offset & (page_size - 1)));
261 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
262 pv->ldp_pages[i], CPT_TRANSIENT);
268 rc = cl_page_own(env, io, clp);
270 LASSERT(clp->cp_state == CPS_FREEING);
271 cl_page_put(env, clp);
277 /* check the page type: if the page is a host page, then do
279 if (clp->cp_type == CPT_CACHEABLE) {
280 struct page *vmpage = cl_page_vmpage(clp);
281 struct page *src_page;
282 struct page *dst_page;
286 src_page = (rw == WRITE) ? pages[i] : vmpage;
287 dst_page = (rw == WRITE) ? vmpage : pages[i];
289 src = ll_kmap_atomic(src_page, KM_USER0);
290 dst = ll_kmap_atomic(dst_page, KM_USER1);
291 memcpy(dst, src, min(page_size, size));
292 ll_kunmap_atomic(dst, KM_USER1);
293 ll_kunmap_atomic(src, KM_USER0);
295 /* make sure page will be added to the transfer by
296 * cl_io_submit()->...->vvp_page_prep_write(). */
298 set_page_dirty(vmpage);
301 /* do not issue the page for read, since it
302 * may reread a ra page which has NOT uptodate
304 cl_page_disown(env, io, clp);
310 cl_2queue_add(queue, clp);
313 * Set page clip to tell transfer formation engine
314 * that page has to be sent even if it is beyond KMS.
316 cl_page_clip(env, clp, 0, min(size, page_size));
321 /* drop the reference count for cl_page_find */
322 cl_page_put(env, clp);
324 file_offset += page_size;
327 if (rc == 0 && io_pages) {
328 rc = cl_io_submit_sync(env, io,
329 rw == READ ? CRT_READ : CRT_WRITE,
335 cl_2queue_discard(env, io, queue);
336 cl_2queue_disown(env, io, queue);
337 cl_2queue_fini(env, queue);
340 EXPORT_SYMBOL(ll_direct_rw_pages);
343 ll_direct_IO_seg(const struct lu_env *env, struct cl_io *io, int rw,
344 struct inode *inode, size_t size, loff_t file_offset,
345 struct page **pages, int page_count)
347 struct ll_dio_pages pvec = { .ldp_pages = pages,
348 .ldp_nr = page_count,
351 .ldp_start_offset = file_offset
354 return ll_direct_rw_pages(env, io, rw, inode, &pvec);
357 #ifdef KMALLOC_MAX_SIZE
358 #define MAX_MALLOC KMALLOC_MAX_SIZE
360 #define MAX_MALLOC (128 * 1024)
363 /* This is the maximum size of a single O_DIRECT request, based on the
364 * kmalloc limit. We need to fit all of the brw_page structs, each one
365 * representing PAGE_SIZE worth of user data, into a single buffer, and
366 * then truncate this to be a full-sized RPC. For 4kB PAGE_SIZE this is
367 * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
368 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
369 ~(DT_MAX_BRW_SIZE - 1))
371 #ifndef HAVE_IOV_ITER_RW
372 # define iov_iter_rw(iter) rw
375 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW)
378 # ifndef HAVE_IOV_ITER_RW
381 struct kiocb *iocb, struct iov_iter *iter,
386 struct file *file = iocb->ki_filp;
387 struct inode *inode = file->f_mapping->host;
388 ssize_t count = iov_iter_count(iter);
389 ssize_t tot_bytes = 0, result = 0;
390 size_t size = MAX_DIO_SIZE;
393 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
394 if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
397 CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
398 "offset=%lld=%llx, pages %zd (max %lu)\n",
399 PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
400 file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
401 MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
403 /* Check that all user buffers are aligned as well */
404 if (iov_iter_alignment(iter) & ~PAGE_MASK)
407 env = cl_env_get(&refcheck);
408 LASSERT(!IS_ERR(env));
409 io = vvp_env_io(env)->vui_cl.cis_io;
412 /* 0. Need locking between buffered and direct access. and race with
413 * size changing by concurrent truncates and writes.
414 * 1. Need inode mutex to operate transient pages.
416 if (iov_iter_rw(iter) == READ)
417 mutex_lock(&inode->i_mutex);
419 while (iov_iter_count(iter)) {
423 count = min_t(size_t, iov_iter_count(iter), size);
424 if (iov_iter_rw(iter) == READ) {
425 if (file_offset >= i_size_read(inode))
428 if (file_offset + count > i_size_read(inode))
429 count = i_size_read(inode) - file_offset;
432 result = iov_iter_get_pages_alloc(iter, &pages, count, &offs);
433 if (likely(result > 0)) {
434 int n = DIV_ROUND_UP(result + offs, PAGE_SIZE);
436 result = ll_direct_IO_seg(env, io, iov_iter_rw(iter),
437 inode, result, file_offset,
439 ll_free_user_pages(pages, n,
440 iov_iter_rw(iter) == READ);
443 if (unlikely(result <= 0)) {
444 /* If we can't allocate a large enough buffer
445 * for the request, shrink it to a smaller
446 * PAGE_SIZE multiple and try again.
447 * We should always be able to kmalloc for a
448 * page worth of page pointers = 4MB on i386. */
449 if (result == -ENOMEM &&
450 size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
452 size = ((((size / 2) - 1) |
453 ~PAGE_MASK) + 1) & PAGE_MASK;
454 CDEBUG(D_VFSTRACE, "DIO size now %zu\n",
462 iov_iter_advance(iter, result);
464 file_offset += result;
467 if (iov_iter_rw(iter) == READ)
468 mutex_unlock(&inode->i_mutex);
471 struct vvp_io *vio = vvp_env_io(env);
473 /* no commit async for direct IO */
474 vio->u.write.vui_written += tot_bytes;
477 cl_env_put(env, &refcheck);
478 return tot_bytes ? : result;
480 #else /* !HAVE_DIRECTIO_ITER && !HAVE_IOV_ITER_RW */
482 ll_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
483 loff_t file_offset, unsigned long nr_segs)
487 struct file *file = iocb->ki_filp;
488 struct inode *inode = file->f_mapping->host;
489 ssize_t count = iov_length(iov, nr_segs);
490 ssize_t tot_bytes = 0, result = 0;
491 unsigned long seg = 0;
492 size_t size = MAX_DIO_SIZE;
496 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
497 if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
500 CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
501 "offset=%lld=%llx, pages %zd (max %lu)\n",
502 PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
503 file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
504 MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
506 /* Check that all user buffers are aligned as well */
507 for (seg = 0; seg < nr_segs; seg++) {
508 if (((unsigned long)iov[seg].iov_base & ~PAGE_MASK) ||
509 (iov[seg].iov_len & ~PAGE_MASK))
513 env = cl_env_get(&refcheck);
514 LASSERT(!IS_ERR(env));
515 io = vvp_env_io(env)->vui_cl.cis_io;
518 for (seg = 0; seg < nr_segs; seg++) {
519 size_t iov_left = iov[seg].iov_len;
520 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
523 if (file_offset >= i_size_read(inode))
525 if (file_offset + iov_left > i_size_read(inode))
526 iov_left = i_size_read(inode) - file_offset;
529 while (iov_left > 0) {
531 int page_count, max_pages = 0;
534 bytes = min(size, iov_left);
535 page_count = ll_get_user_pages(rw, user_addr, bytes,
537 if (likely(page_count > 0)) {
538 if (unlikely(page_count < max_pages))
539 bytes = page_count << PAGE_CACHE_SHIFT;
540 result = ll_direct_IO_seg(env, io, rw, inode,
543 ll_free_user_pages(pages, max_pages, rw==READ);
544 } else if (page_count == 0) {
545 GOTO(out, result = -EFAULT);
549 if (unlikely(result <= 0)) {
550 /* If we can't allocate a large enough buffer
551 * for the request, shrink it to a smaller
552 * PAGE_SIZE multiple and try again.
553 * We should always be able to kmalloc for a
554 * page worth of page pointers = 4MB on i386. */
555 if (result == -ENOMEM &&
556 size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
558 size = ((((size / 2) - 1) |
559 ~PAGE_CACHE_MASK) + 1) &
561 CDEBUG(D_VFSTRACE, "DIO size now %zu\n",
570 file_offset += result;
577 struct vvp_io *vio = vvp_env_io(env);
579 /* no commit async for direct IO */
580 vio->u.write.vui_written += tot_bytes;
583 cl_env_put(env, &refcheck);
584 RETURN(tot_bytes ? tot_bytes : result);
586 #endif /* HAVE_DIRECTIO_ITER || HAVE_IOV_ITER_RW */
589 * Prepare partially written-to page for a write.
591 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
594 struct cl_attr *attr = vvp_env_thread_attr(env);
595 struct cl_object *obj = io->ci_obj;
596 struct vvp_page *vpg = cl_object_page_slice(obj, pg);
597 loff_t offset = cl_offset(obj, vvp_index(vpg));
600 cl_object_attr_lock(obj);
601 result = cl_object_attr_get(env, obj, attr);
602 cl_object_attr_unlock(obj);
605 * If are writing to a new page, no need to read old data.
606 * The extent locking will have updated the KMS, and for our
607 * purposes here we can treat it like i_size.
609 if (attr->cat_kms <= offset) {
610 char *kaddr = ll_kmap_atomic(vpg->vpg_page, KM_USER0);
612 memset(kaddr, 0, cl_page_size(obj));
613 ll_kunmap_atomic(kaddr, KM_USER0);
614 } else if (vpg->vpg_defer_uptodate)
615 vpg->vpg_ra_used = 1;
617 result = ll_page_sync_io(env, io, pg, CRT_READ);
622 static int ll_write_begin(struct file *file, struct address_space *mapping,
623 loff_t pos, unsigned len, unsigned flags,
624 struct page **pagep, void **fsdata)
626 struct ll_cl_context *lcc;
627 const struct lu_env *env;
629 struct cl_page *page;
631 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
632 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
633 struct page *vmpage = NULL;
634 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
635 unsigned to = from + len;
639 CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
641 lcc = ll_cl_find(file);
643 GOTO(out, result = -EIO);
648 /* To avoid deadlock, try to lock page first. */
649 vmpage = grab_cache_page_nowait(mapping, index);
651 if (unlikely(vmpage == NULL ||
652 PageDirty(vmpage) || PageWriteback(vmpage))) {
653 struct vvp_io *vio = vvp_env_io(env);
654 struct cl_page_list *plist = &vio->u.write.vui_queue;
656 /* if the page is already in dirty cache, we have to commit
657 * the pages right now; otherwise, it may cause deadlock
658 * because it holds page lock of a dirty page and request for
659 * more grants. It's okay for the dirty page to be the first
660 * one in commit page list, though. */
661 if (vmpage != NULL && plist->pl_nr > 0) {
663 page_cache_release(vmpage);
667 /* commit pages and then wait for page lock */
668 result = vvp_io_write_commit(env, io);
672 if (vmpage == NULL) {
673 vmpage = grab_cache_page_write_begin(mapping, index,
676 GOTO(out, result = -ENOMEM);
680 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
682 GOTO(out, result = PTR_ERR(page));
684 lcc->lcc_page = page;
685 lu_ref_add(&page->cp_reference, "cl_io", io);
687 cl_page_assume(env, io, page);
688 if (!PageUptodate(vmpage)) {
690 * We're completely overwriting an existing page,
691 * so _don't_ set it up to date until commit_write
693 if (from == 0 && to == PAGE_SIZE) {
694 CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
695 POISON_PAGE(vmpage, 0x11);
697 /* TODO: can be optimized at OSC layer to check if it
698 * is a lockless IO. In that case, it's not necessary
699 * to read the data. */
700 result = ll_prepare_partial_page(env, io, page);
702 SetPageUptodate(vmpage);
706 cl_page_unassume(env, io, page);
710 if (vmpage != NULL) {
712 page_cache_release(vmpage);
721 static int ll_write_end(struct file *file, struct address_space *mapping,
722 loff_t pos, unsigned len, unsigned copied,
723 struct page *vmpage, void *fsdata)
725 struct ll_cl_context *lcc = fsdata;
726 const struct lu_env *env;
729 struct cl_page *page;
730 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
735 page_cache_release(vmpage);
737 LASSERT(lcc != NULL);
739 page = lcc->lcc_page;
741 vio = vvp_env_io(env);
743 LASSERT(cl_page_is_owned(page, io));
745 struct cl_page_list *plist = &vio->u.write.vui_queue;
747 lcc->lcc_page = NULL; /* page will be queued */
749 /* Add it into write queue */
750 cl_page_list_add(plist, page);
751 if (plist->pl_nr == 1) /* first page */
752 vio->u.write.vui_from = from;
755 vio->u.write.vui_to = from + copied;
757 /* To address the deadlock in balance_dirty_pages() where
758 * this dirty page may be written back in the same thread. */
759 if (PageDirty(vmpage))
762 /* We may have one full RPC, commit it soon */
763 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
766 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
767 "queued page: %d.\n", plist->pl_nr);
769 cl_page_disown(env, io, page);
771 lcc->lcc_page = NULL;
772 lu_ref_del(&page->cp_reference, "cl_io", io);
773 cl_page_put(env, page);
775 /* page list is not contiguous now, commit it now */
779 file->f_flags & O_SYNC || IS_SYNC(file->f_path.dentry->d_inode))
780 result = vvp_io_write_commit(env, io);
782 RETURN(result >= 0 ? copied : result);
785 #ifdef CONFIG_MIGRATION
786 static int ll_migratepage(struct address_space *mapping,
787 struct page *newpage, struct page *page
788 #ifdef HAVE_MIGRATEPAGE_4ARGS
789 , enum migrate_mode mode
793 /* Always fail page migration until we have a proper implementation */
798 #ifndef MS_HAS_NEW_AOPS
799 const struct address_space_operations ll_aops = {
800 .readpage = ll_readpage,
801 .direct_IO = ll_direct_IO,
802 .writepage = ll_writepage,
803 .writepages = ll_writepages,
804 .set_page_dirty = __set_page_dirty_nobuffers,
805 .write_begin = ll_write_begin,
806 .write_end = ll_write_end,
807 .invalidatepage = ll_invalidatepage,
808 .releasepage = (void *)ll_releasepage,
809 #ifdef CONFIG_MIGRATION
810 .migratepage = ll_migratepage,
814 const struct address_space_operations_ext ll_aops = {
815 .orig_aops.readpage = ll_readpage,
816 .orig_aops.direct_IO = ll_direct_IO,
817 .orig_aops.writepage = ll_writepage,
818 .orig_aops.writepages = ll_writepages,
819 .orig_aops.set_page_dirty = __set_page_dirty_nobuffers,
820 .orig_aops.invalidatepage = ll_invalidatepage,
821 .orig_aops.releasepage = ll_releasepage,
822 #ifdef CONFIG_MIGRATION
823 .orig_aops.migratepage = ll_migratepage,
825 .write_begin = ll_write_begin,
826 .write_end = ll_write_end