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)
131 struct cl_object *obj;
132 struct cl_page *page;
133 struct address_space *mapping;
136 LASSERT(PageLocked(vmpage));
137 if (PageWriteback(vmpage) || PageDirty(vmpage))
140 mapping = vmpage->mapping;
144 obj = ll_i2info(mapping->host)->lli_clob;
148 /* 1 for caller, 1 for cl_page and 1 for page cache */
149 if (page_count(vmpage) > 3)
152 page = cl_vmpage_page(vmpage, obj);
156 env = cl_env_percpu_get();
157 LASSERT(!IS_ERR(env));
159 if (!cl_page_in_use(page)) {
161 cl_page_delete(env, page);
164 /* To use percpu env array, the call path can not be rescheduled;
165 * otherwise percpu array will be messed if ll_releaspage() called
166 * again on the same CPU.
168 * If this page holds the last refc of cl_object, the following
169 * call path may cause reschedule:
170 * cl_page_put -> cl_page_free -> cl_object_put ->
171 * lu_object_put -> lu_object_free -> lov_delete_raid0.
173 * However, the kernel can't get rid of this inode until all pages have
174 * been cleaned up. Now that we hold page lock here, it's pretty safe
175 * that we won't get into object delete path.
177 LASSERT(cl_object_refc(obj) > 1);
178 cl_page_put(env, page);
180 cl_env_percpu_put(env);
184 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
186 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
187 int rw, struct inode *inode,
188 struct ll_dio_pages *pv)
191 struct cl_2queue *queue;
192 struct cl_object *obj = io->ci_obj;
195 loff_t file_offset = pv->ldp_start_offset;
196 size_t size = pv->ldp_size;
197 int page_count = pv->ldp_nr;
198 struct page **pages = pv->ldp_pages;
199 size_t page_size = cl_page_size(obj);
204 queue = &io->ci_queue;
205 cl_2queue_init(queue);
206 for (i = 0; i < page_count; i++) {
208 file_offset = pv->ldp_offsets[i];
210 LASSERT(!(file_offset & (page_size - 1)));
211 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
212 pv->ldp_pages[i], CPT_TRANSIENT);
218 rc = cl_page_own(env, io, clp);
220 LASSERT(clp->cp_state == CPS_FREEING);
221 cl_page_put(env, clp);
227 /* check the page type: if the page is a host page, then do
229 if (clp->cp_type == CPT_CACHEABLE) {
230 struct page *vmpage = cl_page_vmpage(clp);
231 struct page *src_page;
232 struct page *dst_page;
236 src_page = (rw == WRITE) ? pages[i] : vmpage;
237 dst_page = (rw == WRITE) ? vmpage : pages[i];
239 src = ll_kmap_atomic(src_page, KM_USER0);
240 dst = ll_kmap_atomic(dst_page, KM_USER1);
241 memcpy(dst, src, min(page_size, size));
242 ll_kunmap_atomic(dst, KM_USER1);
243 ll_kunmap_atomic(src, KM_USER0);
245 /* make sure page will be added to the transfer by
246 * cl_io_submit()->...->vvp_page_prep_write(). */
248 set_page_dirty(vmpage);
251 /* do not issue the page for read, since it
252 * may reread a ra page which has NOT uptodate
254 cl_page_disown(env, io, clp);
260 cl_2queue_add(queue, clp);
263 * Set page clip to tell transfer formation engine
264 * that page has to be sent even if it is beyond KMS.
266 cl_page_clip(env, clp, 0, min(size, page_size));
271 /* drop the reference count for cl_page_find */
272 cl_page_put(env, clp);
274 file_offset += page_size;
277 if (rc == 0 && io_pages) {
278 rc = cl_io_submit_sync(env, io,
279 rw == READ ? CRT_READ : CRT_WRITE,
285 cl_2queue_discard(env, io, queue);
286 cl_2queue_disown(env, io, queue);
287 cl_2queue_fini(env, queue);
290 EXPORT_SYMBOL(ll_direct_rw_pages);
293 ll_direct_IO_seg(const struct lu_env *env, struct cl_io *io, int rw,
294 struct inode *inode, size_t size, loff_t file_offset,
295 struct page **pages, int page_count)
297 struct ll_dio_pages pvec = { .ldp_pages = pages,
298 .ldp_nr = page_count,
301 .ldp_start_offset = file_offset
304 return ll_direct_rw_pages(env, io, rw, inode, &pvec);
307 /* ll_free_user_pages - tear down page struct array
308 * @pages: array of page struct pointers underlying target buffer */
309 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
313 for (i = 0; i < npages; i++) {
314 if (pages[i] == NULL)
317 set_page_dirty_lock(pages[i]);
318 page_cache_release(pages[i]);
321 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW)
324 OBD_FREE_LARGE(pages, npages * sizeof(*pages));
328 #ifdef KMALLOC_MAX_SIZE
329 #define MAX_MALLOC KMALLOC_MAX_SIZE
331 #define MAX_MALLOC (128 * 1024)
334 /* This is the maximum size of a single O_DIRECT request, based on the
335 * kmalloc limit. We need to fit all of the brw_page structs, each one
336 * representing PAGE_SIZE worth of user data, into a single buffer, and
337 * then truncate this to be a full-sized RPC. For 4kB PAGE_SIZE this is
338 * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
339 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
340 ~(DT_MAX_BRW_SIZE - 1))
342 #ifndef HAVE_IOV_ITER_RW
343 # define iov_iter_rw(iter) rw
346 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW)
349 # ifndef HAVE_IOV_ITER_RW
352 struct kiocb *iocb, struct iov_iter *iter,
355 struct ll_cl_context *lcc;
356 const struct lu_env *env;
358 struct file *file = iocb->ki_filp;
359 struct inode *inode = file->f_mapping->host;
360 ssize_t count = iov_iter_count(iter);
361 ssize_t tot_bytes = 0, result = 0;
362 size_t size = MAX_DIO_SIZE;
364 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
365 if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
368 CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
369 "offset=%lld=%llx, pages %zd (max %lu)\n",
370 PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
371 file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
372 MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
374 /* Check that all user buffers are aligned as well */
375 if (iov_iter_alignment(iter) & ~PAGE_MASK)
378 lcc = ll_cl_find(file);
383 LASSERT(!IS_ERR(env));
387 /* 0. Need locking between buffered and direct access. and race with
388 * size changing by concurrent truncates and writes.
389 * 1. Need inode mutex to operate transient pages.
391 if (iov_iter_rw(iter) == READ)
392 mutex_lock(&inode->i_mutex);
394 while (iov_iter_count(iter)) {
398 count = min_t(size_t, iov_iter_count(iter), size);
399 if (iov_iter_rw(iter) == READ) {
400 if (file_offset >= i_size_read(inode))
403 if (file_offset + count > i_size_read(inode))
404 count = i_size_read(inode) - file_offset;
407 result = iov_iter_get_pages_alloc(iter, &pages, count, &offs);
408 if (likely(result > 0)) {
409 int n = DIV_ROUND_UP(result + offs, PAGE_SIZE);
411 result = ll_direct_IO_seg(env, io, iov_iter_rw(iter),
412 inode, result, file_offset,
414 ll_free_user_pages(pages, n,
415 iov_iter_rw(iter) == READ);
418 if (unlikely(result <= 0)) {
419 /* If we can't allocate a large enough buffer
420 * for the request, shrink it to a smaller
421 * PAGE_SIZE multiple and try again.
422 * We should always be able to kmalloc for a
423 * page worth of page pointers = 4MB on i386. */
424 if (result == -ENOMEM &&
425 size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
427 size = ((((size / 2) - 1) |
428 ~PAGE_MASK) + 1) & PAGE_MASK;
429 CDEBUG(D_VFSTRACE, "DIO size now %zu\n",
437 iov_iter_advance(iter, result);
439 file_offset += result;
442 if (iov_iter_rw(iter) == READ)
443 mutex_unlock(&inode->i_mutex);
446 struct vvp_io *vio = vvp_env_io(env);
448 /* no commit async for direct IO */
449 vio->u.write.vui_written += tot_bytes;
452 return tot_bytes ? : result;
454 #else /* !HAVE_DIRECTIO_ITER && !HAVE_IOV_ITER_RW */
456 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
457 size_t size, struct page ***pages,
460 int result = -ENOMEM;
462 /* set an arbitrary limit to prevent arithmetic overflow */
463 if (size > MAX_DIRECTIO_SIZE) {
468 *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >>
470 *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
472 OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
474 down_read(¤t->mm->mmap_sem);
475 result = get_user_pages(current, current->mm, user_addr,
476 *max_pages, (rw == READ), 0, *pages,
478 up_read(¤t->mm->mmap_sem);
479 if (unlikely(result <= 0))
480 OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
487 ll_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
488 loff_t file_offset, unsigned long nr_segs)
490 struct ll_cl_context *lcc;
491 const struct lu_env *env;
493 struct file *file = iocb->ki_filp;
494 struct inode *inode = file->f_mapping->host;
495 ssize_t count = iov_length(iov, nr_segs);
496 ssize_t tot_bytes = 0, result = 0;
497 unsigned long seg = 0;
498 size_t size = MAX_DIO_SIZE;
501 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
502 if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
505 CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
506 "offset=%lld=%llx, pages %zd (max %lu)\n",
507 PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
508 file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
509 MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
511 /* Check that all user buffers are aligned as well */
512 for (seg = 0; seg < nr_segs; seg++) {
513 if (((unsigned long)iov[seg].iov_base & ~PAGE_MASK) ||
514 (iov[seg].iov_len & ~PAGE_MASK))
518 lcc = ll_cl_find(file);
523 LASSERT(!IS_ERR(env));
527 for (seg = 0; seg < nr_segs; seg++) {
528 size_t iov_left = iov[seg].iov_len;
529 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
532 if (file_offset >= i_size_read(inode))
534 if (file_offset + iov_left > i_size_read(inode))
535 iov_left = i_size_read(inode) - file_offset;
538 while (iov_left > 0) {
540 int page_count, max_pages = 0;
543 bytes = min(size, iov_left);
544 page_count = ll_get_user_pages(rw, user_addr, bytes,
546 if (likely(page_count > 0)) {
547 if (unlikely(page_count < max_pages))
548 bytes = page_count << PAGE_CACHE_SHIFT;
549 result = ll_direct_IO_seg(env, io, rw, inode,
552 ll_free_user_pages(pages, max_pages, rw==READ);
553 } else if (page_count == 0) {
554 GOTO(out, result = -EFAULT);
558 if (unlikely(result <= 0)) {
559 /* If we can't allocate a large enough buffer
560 * for the request, shrink it to a smaller
561 * PAGE_SIZE multiple and try again.
562 * We should always be able to kmalloc for a
563 * page worth of page pointers = 4MB on i386. */
564 if (result == -ENOMEM &&
565 size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
567 size = ((((size / 2) - 1) |
568 ~PAGE_CACHE_MASK) + 1) &
570 CDEBUG(D_VFSTRACE, "DIO size now %zu\n",
579 file_offset += result;
586 struct vvp_io *vio = vvp_env_io(env);
588 /* no commit async for direct IO */
589 vio->u.write.vui_written += tot_bytes;
592 RETURN(tot_bytes ? tot_bytes : result);
594 #endif /* HAVE_DIRECTIO_ITER || HAVE_IOV_ITER_RW */
597 * Prepare partially written-to page for a write.
599 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
602 struct cl_attr *attr = vvp_env_thread_attr(env);
603 struct cl_object *obj = io->ci_obj;
604 struct vvp_page *vpg = cl_object_page_slice(obj, pg);
605 loff_t offset = cl_offset(obj, vvp_index(vpg));
608 cl_object_attr_lock(obj);
609 result = cl_object_attr_get(env, obj, attr);
610 cl_object_attr_unlock(obj);
613 * If are writing to a new page, no need to read old data.
614 * The extent locking will have updated the KMS, and for our
615 * purposes here we can treat it like i_size.
617 if (attr->cat_kms <= offset) {
618 char *kaddr = ll_kmap_atomic(vpg->vpg_page, KM_USER0);
620 memset(kaddr, 0, cl_page_size(obj));
621 ll_kunmap_atomic(kaddr, KM_USER0);
622 } else if (vpg->vpg_defer_uptodate)
623 vpg->vpg_ra_used = 1;
625 result = ll_page_sync_io(env, io, pg, CRT_READ);
630 static int ll_write_begin(struct file *file, struct address_space *mapping,
631 loff_t pos, unsigned len, unsigned flags,
632 struct page **pagep, void **fsdata)
634 struct ll_cl_context *lcc;
635 const struct lu_env *env;
637 struct cl_page *page;
639 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
640 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
641 struct page *vmpage = NULL;
642 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
643 unsigned to = from + len;
647 CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
649 lcc = ll_cl_find(file);
652 GOTO(out, result = -EIO);
658 /* To avoid deadlock, try to lock page first. */
659 vmpage = grab_cache_page_nowait(mapping, index);
661 if (unlikely(vmpage == NULL ||
662 PageDirty(vmpage) || PageWriteback(vmpage))) {
663 struct vvp_io *vio = vvp_env_io(env);
664 struct cl_page_list *plist = &vio->u.write.vui_queue;
666 /* if the page is already in dirty cache, we have to commit
667 * the pages right now; otherwise, it may cause deadlock
668 * because it holds page lock of a dirty page and request for
669 * more grants. It's okay for the dirty page to be the first
670 * one in commit page list, though. */
671 if (vmpage != NULL && plist->pl_nr > 0) {
673 page_cache_release(vmpage);
677 /* commit pages and then wait for page lock */
678 result = vvp_io_write_commit(env, io);
682 if (vmpage == NULL) {
683 vmpage = grab_cache_page_write_begin(mapping, index,
686 GOTO(out, result = -ENOMEM);
690 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
692 GOTO(out, result = PTR_ERR(page));
694 lcc->lcc_page = page;
695 lu_ref_add(&page->cp_reference, "cl_io", io);
697 cl_page_assume(env, io, page);
698 if (!PageUptodate(vmpage)) {
700 * We're completely overwriting an existing page,
701 * so _don't_ set it up to date until commit_write
703 if (from == 0 && to == PAGE_SIZE) {
704 CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
705 POISON_PAGE(vmpage, 0x11);
707 /* TODO: can be optimized at OSC layer to check if it
708 * is a lockless IO. In that case, it's not necessary
709 * to read the data. */
710 result = ll_prepare_partial_page(env, io, page);
712 SetPageUptodate(vmpage);
716 cl_page_unassume(env, io, page);
720 if (vmpage != NULL) {
722 page_cache_release(vmpage);
725 io->ci_result = result;
733 static int ll_write_end(struct file *file, struct address_space *mapping,
734 loff_t pos, unsigned len, unsigned copied,
735 struct page *vmpage, void *fsdata)
737 struct ll_cl_context *lcc = fsdata;
738 const struct lu_env *env;
741 struct cl_page *page;
742 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
747 page_cache_release(vmpage);
749 LASSERT(lcc != NULL);
751 page = lcc->lcc_page;
753 vio = vvp_env_io(env);
755 LASSERT(cl_page_is_owned(page, io));
757 struct cl_page_list *plist = &vio->u.write.vui_queue;
759 lcc->lcc_page = NULL; /* page will be queued */
761 /* Add it into write queue */
762 cl_page_list_add(plist, page);
763 if (plist->pl_nr == 1) /* first page */
764 vio->u.write.vui_from = from;
767 vio->u.write.vui_to = from + copied;
769 /* To address the deadlock in balance_dirty_pages() where
770 * this dirty page may be written back in the same thread. */
771 if (PageDirty(vmpage))
774 /* We may have one full RPC, commit it soon */
775 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
778 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
779 "queued page: %d.\n", plist->pl_nr);
781 cl_page_disown(env, io, page);
783 lcc->lcc_page = NULL;
784 lu_ref_del(&page->cp_reference, "cl_io", io);
785 cl_page_put(env, page);
787 /* page list is not contiguous now, commit it now */
791 file->f_flags & O_SYNC || IS_SYNC(file->f_path.dentry->d_inode))
792 result = vvp_io_write_commit(env, io);
795 io->ci_result = result;
796 RETURN(result >= 0 ? copied : result);
799 #ifdef CONFIG_MIGRATION
800 static int ll_migratepage(struct address_space *mapping,
801 struct page *newpage, struct page *page
802 #ifdef HAVE_MIGRATEPAGE_4ARGS
803 , enum migrate_mode mode
807 /* Always fail page migration until we have a proper implementation */
812 const struct address_space_operations ll_aops = {
813 .readpage = ll_readpage,
814 .direct_IO = ll_direct_IO,
815 .writepage = ll_writepage,
816 .writepages = ll_writepages,
817 .set_page_dirty = __set_page_dirty_nobuffers,
818 .write_begin = ll_write_begin,
819 .write_end = ll_write_end,
820 .invalidatepage = ll_invalidatepage,
821 .releasepage = (void *)ll_releasepage,
822 #ifdef CONFIG_MIGRATION
823 .migratepage = ll_migratepage,